Skip to content

horde_sdk.generation_parameters.image.constraints

Per-sampler capability and constraint data shared by clients, the API and workers.

The image backend accepts a different set of solver knobs for every sampler, and rejects the rest with a type error raised from inside the render graph. This module states, once, which knobs each sampler takes, over what range, what a step of it costs, and which sampler/scheduler/baseline combinations are known to produce no usable image. Requests can then be checked before anything is dispatched.

Critical public members:

Every fact here is separated into two kinds. Hard constraints and knob applicability are read out of the backend's own source: a knob is applicable when the solver function accepts a keyword of that name, and a numeric range is the one the backend's own node for that sampler declares. Recommendations are opinions, and each carries a CONSTRAINT_PROVENANCE saying whose opinion it is.

This module imports only from consts and the model reference, so the wire models and the worker feature flags can both depend on it.

MEASURED_COST_RATIO_PROVENANCE module-attribute 

MEASURED_COST_RATIO_PROVENANCE = (
    CONSTRAINT_PROVENANCE.measured
)

The provenance both measured cost ratios carry.

The figures come from MEASURED_COST_RATIO_SOURCE: renders taken through the production image pipeline on one card, at step counts 10, 20, 30 and 40, three repeats each, warmups discarded, with the seed, prompt and cfg scale held fixed. Each ratio is the slope of an ordinary least-squares fit of median wall time against step count, divided by k_euler's slope on the same model. Fitting the slope is what separates per-step cost from the fixed per-render overhead, which is around 0.29s at 512x512 and 1.40s at 1024x1024.

The measurement corroborates the fixed-rate sampler work profiles: at 1024x1024 every fixed sampler lands close to its marginal work family. These ratios are one card's evidence, not a portable price or an execution ceiling.

MEASURED_COST_RATIO_SOURCE module-attribute 

MEASURED_COST_RATIO_SOURCE = (
    "sampler-cost-2026-08-03T23-02-56.165788Z.json"
)

Filename of the measurement artifact both ratio columns were read from.

Published alongside the other parameter-sweep measurements, and carrying every timing sample, fit coefficient and setting behind the two rounded numbers held here.

SAMPLER_CONSTRAINTS module-attribute 

SAMPLER_CONSTRAINTS: Mapping[
    KNOWN_IMAGE_SAMPLERS, SamplerConstraints
] = MappingProxyType(
    dict(
        (
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.k_lms,
                backend_solver_function="sample_lms",
                numeric_knob_ranges=MappingProxyType(
                    {
                        SAMPLER_SOLVER_KNOB.order: _LMS_ORDER_RANGE
                    }
                ),
                measured_cost_ratio_sd15=1.02,
                measured_cost_ratio_sdxl=1.0,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.k_heun,
                backend_solver_function="sample_heun",
                numeric_knob_ranges=_CHURN_KNOBS,
                measured_cost_ratio_sd15=2.03,
                measured_cost_ratio_sdxl=1.87,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.k_euler,
                backend_solver_function="sample_euler",
                numeric_knob_ranges=_CHURN_KNOBS,
                measured_cost_ratio_sd15=1.0,
                measured_cost_ratio_sdxl=1.0,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.k_euler_a,
                backend_solver_function="sample_euler_ancestral",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                measured_cost_ratio_sd15=1.08,
                measured_cost_ratio_sdxl=1.01,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.k_dpm_2,
                backend_solver_function="sample_dpm_2",
                numeric_knob_ranges=_CHURN_KNOBS,
                measured_cost_ratio_sd15=2.17,
                measured_cost_ratio_sdxl=1.94,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.k_dpm_2_a,
                backend_solver_function="sample_dpm_2_ancestral",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                measured_cost_ratio_sd15=2.22,
                measured_cost_ratio_sdxl=1.97,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.k_dpm_fast,
                backend_solver_function="sample_dpm_fast",
                measured_cost_ratio_sd15=1.16,
                measured_cost_ratio_sdxl=0.97,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.k_dpm_adaptive,
                backend_solver_function="sample_dpm_adaptive",
                numeric_knob_ranges=MappingProxyType(
                    {
                        SAMPLER_SOLVER_KNOB.eta: _ETA_RANGE_OFF_BY_DEFAULT,
                        SAMPLER_SOLVER_KNOB.s_noise: _S_NOISE_RANGE,
                        SAMPLER_SOLVER_KNOB.order: _DPM_ADAPTIVE_ORDER_RANGE,
                    }
                ),
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.k_dpmpp_2s_a,
                backend_solver_function="sample_dpmpp_2s_ancestral",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                measured_cost_ratio_sd15=2.35,
                measured_cost_ratio_sdxl=2.01,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.k_dpmpp_2m,
                backend_solver_function="sample_dpmpp_2m",
                measured_cost_ratio_sd15=1.08,
                measured_cost_ratio_sdxl=0.98,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.dpmsolver,
                backend_solver_function="sample_dpmpp_2m",
                measured_cost_ratio_sd15=1.12,
                measured_cost_ratio_sdxl=1.0,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.k_dpmpp_sde,
                backend_solver_function="sample_dpmpp_sde",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                measured_cost_ratio_sd15=2.58,
                measured_cost_ratio_sdxl=2.22,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.lcm,
                backend_solver_function="sample_lcm",
                numeric_knob_ranges=_S_NOISE_ONLY,
                measured_cost_ratio_sd15=1.21,
                measured_cost_ratio_sdxl=0.96,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.DDIM,
                backend_solver_function="sample_euler",
                measured_cost_ratio_sd15=1.13,
                measured_cost_ratio_sdxl=0.96,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.uni_pc,
                backend_solver_function="sample_unipc",
                measured_cost_ratio_sd15=1.34,
                measured_cost_ratio_sdxl=0.98,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.uni_pc_bh2,
                backend_solver_function="sample_unipc_bh2",
                measured_cost_ratio_sd15=1.17,
                measured_cost_ratio_sdxl=0.99,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.dpmpp_2m_sde,
                backend_solver_function="sample_dpmpp_2m_sde",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                solver_type_choices=_MIDPOINT_OR_HEUN,
                measured_cost_ratio_sd15=1.33,
                measured_cost_ratio_sdxl=1.14,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.dpmpp_3m_sde,
                backend_solver_function="sample_dpmpp_3m_sde",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                measured_cost_ratio_sd15=1.18,
                measured_cost_ratio_sdxl=1.12,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.ddpm,
                backend_solver_function="sample_ddpm",
                measured_cost_ratio_sd15=1.05,
                measured_cost_ratio_sdxl=1.03,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.deis,
                backend_solver_function="sample_deis",
                numeric_knob_ranges=MappingProxyType(
                    {
                        SAMPLER_SOLVER_KNOB.order: _DEIS_MAX_ORDER_RANGE
                    }
                ),
                measured_cost_ratio_sd15=1.13,
                measured_cost_ratio_sdxl=0.96,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.ipndm,
                backend_solver_function="sample_ipndm",
                numeric_knob_ranges=MappingProxyType(
                    {
                        SAMPLER_SOLVER_KNOB.order: _IPNDM_MAX_ORDER_RANGE
                    }
                ),
                measured_cost_ratio_sd15=1.01,
                measured_cost_ratio_sdxl=1.0,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.res_multistep,
                backend_solver_function="sample_res_multistep",
                numeric_knob_ranges=_S_NOISE_ONLY,
                measured_cost_ratio_sd15=1.01,
                measured_cost_ratio_sdxl=0.98,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.gradient_estimation,
                backend_solver_function="sample_gradient_estimation",
                measured_cost_ratio_sd15=1.18,
                measured_cost_ratio_sdxl=0.99,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.heunpp2,
                backend_solver_function="sample_heunpp2",
                numeric_knob_ranges=_CHURN_KNOBS,
                measured_cost_ratio_sd15=3.21,
                measured_cost_ratio_sdxl=2.93,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.er_sde,
                backend_solver_function="sample_er_sde",
                numeric_knob_ranges=_S_NOISE_ONLY,
                measured_cost_ratio_sd15=1.07,
                measured_cost_ratio_sdxl=0.98,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.sa_solver,
                backend_solver_function="sample_sa_solver",
                numeric_knob_ranges=_S_NOISE_ONLY,
                measured_cost_ratio_sd15=1.08,
                measured_cost_ratio_sdxl=0.94,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.euler_cfg_pp,
                backend_solver_function="sample_euler_cfg_pp",
                measured_cost_ratio_sd15=1.11,
                measured_cost_ratio_sdxl=1.1,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.euler_ancestral_cfg_pp,
                backend_solver_function="sample_euler_ancestral_cfg_pp",
                numeric_knob_ranges=MappingProxyType(
                    {
                        SAMPLER_SOLVER_KNOB.eta: _ETA_RANGE_UNIT,
                        SAMPLER_SOLVER_KNOB.s_noise: _S_NOISE_RANGE_NARROW,
                    }
                ),
                measured_cost_ratio_sd15=1.15,
                measured_cost_ratio_sdxl=0.99,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.exp_heun_2_x0,
                backend_solver_function="sample_exp_heun_2_x0",
                solver_type_choices=_PHI_1_OR_PHI_2,
                measured_cost_ratio_sd15=2.08,
                measured_cost_ratio_sdxl=2.39,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.exp_heun_2_x0_sde,
                backend_solver_function="sample_exp_heun_2_x0_sde",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                solver_type_choices=_PHI_1_OR_PHI_2,
                measured_cost_ratio_sd15=2.27,
                measured_cost_ratio_sdxl=1.99,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.dpmpp_2s_ancestral_cfg_pp,
                backend_solver_function="sample_dpmpp_2s_ancestral_cfg_pp",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                measured_cost_ratio_sd15=2.14,
                measured_cost_ratio_sdxl=2.02,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.dpmpp_2m_cfg_pp,
                backend_solver_function="sample_dpmpp_2m_cfg_pp",
                measured_cost_ratio_sd15=1.01,
                measured_cost_ratio_sdxl=1.05,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.dpmpp_2m_sde_heun,
                backend_solver_function="sample_dpmpp_2m_sde_heun",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                solver_type_choices=_MIDPOINT_OR_HEUN,
                measured_cost_ratio_sd15=1.27,
                measured_cost_ratio_sdxl=1.1,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.ipndm_v,
                backend_solver_function="sample_ipndm_v",
                numeric_knob_ranges=MappingProxyType(
                    {
                        SAMPLER_SOLVER_KNOB.order: _IPNDM_MAX_ORDER_RANGE
                    }
                ),
                measured_cost_ratio_sd15=1.11,
                measured_cost_ratio_sdxl=1.15,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.res_multistep_cfg_pp,
                backend_solver_function="sample_res_multistep_cfg_pp",
                numeric_knob_ranges=_S_NOISE_ONLY,
                measured_cost_ratio_sd15=1.08,
                measured_cost_ratio_sdxl=1.03,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.res_multistep_ancestral,
                backend_solver_function="sample_res_multistep_ancestral",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                measured_cost_ratio_sd15=1.13,
                measured_cost_ratio_sdxl=1.14,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.res_multistep_ancestral_cfg_pp,
                backend_solver_function="sample_res_multistep_ancestral_cfg_pp",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                measured_cost_ratio_sd15=1.11,
                measured_cost_ratio_sdxl=1.04,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.gradient_estimation_cfg_pp,
                backend_solver_function="sample_gradient_estimation_cfg_pp",
                measured_cost_ratio_sd15=1.2,
                measured_cost_ratio_sdxl=1.02,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.seeds_2,
                backend_solver_function="sample_seeds_2",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                solver_type_choices=_PHI_1_OR_PHI_2,
                measured_cost_ratio_sd15=2.26,
                measured_cost_ratio_sdxl=1.88,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.seeds_3,
                backend_solver_function="sample_seeds_3",
                numeric_knob_ranges=_ANCESTRAL_KNOBS,
                measured_cost_ratio_sd15=3.52,
                measured_cost_ratio_sdxl=2.82,
            ),
            _constraints(
                sampler=KNOWN_IMAGE_SAMPLERS.sa_solver_pece,
                backend_solver_function="sample_sa_solver_pece",
                numeric_knob_ranges=_S_NOISE_ONLY,
                measured_cost_ratio_sd15=2.38,
                measured_cost_ratio_sdxl=2.11,
            ),
        )
    )
)

One record per known sampler, covering knob applicability, ranges, cost and identity.

REJECTED_SAMPLER_SCHEDULER_PAIRINGS module-attribute 

REJECTED_SAMPLER_SCHEDULER_PAIRINGS: frozenset[
    tuple[KNOWN_IMAGE_SAMPLERS, KNOWN_IMAGE_SCHEDULERS]
] = frozenset(
    {
        (
            KNOWN_IMAGE_SAMPLERS.dpmpp_3m_sde,
            KNOWN_IMAGE_SCHEDULERS.normal,
        )
    }
)

Sampler and scheduler pairings that produce no usable image and are refused rather than substituted.

SCHEDULER_BASELINE_APPLICABILITY module-attribute 

SCHEDULER_BASELINE_APPLICABILITY: Mapping[
    KNOWN_IMAGE_SCHEDULERS,
    frozenset[KNOWN_IMAGE_GENERATION_BASELINE],
] = MappingProxyType(
    {
        KNOWN_IMAGE_SCHEDULERS.align_your_steps: frozenset(
            {
                KNOWN_IMAGE_GENERATION_BASELINE.stable_diffusion_1,
                KNOWN_IMAGE_GENERATION_BASELINE.stable_diffusion_xl,
            }
        ),
        KNOWN_IMAGE_SCHEDULERS.gits: frozenset(
            {
                KNOWN_IMAGE_GENERATION_BASELINE.stable_diffusion_1,
                KNOWN_IMAGE_GENERATION_BASELINE.stable_diffusion_xl,
            }
        ),
    }
)

Baselines each restricted scheduler is defined for. A scheduler absent here works on every baseline.

CFG_PP_SAMPLERS module-attribute 

CFG_PP_SAMPLERS: frozenset[KNOWN_IMAGE_SAMPLERS] = (
    frozenset(
        {
            KNOWN_IMAGE_SAMPLERS.euler_cfg_pp,
            KNOWN_IMAGE_SAMPLERS.euler_ancestral_cfg_pp,
            KNOWN_IMAGE_SAMPLERS.dpmpp_2s_ancestral_cfg_pp,
            KNOWN_IMAGE_SAMPLERS.dpmpp_2m_cfg_pp,
            KNOWN_IMAGE_SAMPLERS.res_multistep_cfg_pp,
            KNOWN_IMAGE_SAMPLERS.res_multistep_ancestral_cfg_pp,
            KNOWN_IMAGE_SAMPLERS.gradient_estimation_cfg_pp,
        }
    )
)

The samplers applying the CFG++ correction, which expect a much lower cfg_scale than the rest.

RECOMMENDED_SAMPLERS module-attribute 

RECOMMENDED_SAMPLERS: frozenset[KNOWN_IMAGE_SAMPLERS] = (
    frozenset(
        {
            KNOWN_IMAGE_SAMPLERS.k_euler,
            KNOWN_IMAGE_SAMPLERS.k_euler_a,
            KNOWN_IMAGE_SAMPLERS.k_dpmpp_2m,
            KNOWN_IMAGE_SAMPLERS.dpmpp_2m_sde,
            KNOWN_IMAGE_SAMPLERS.k_dpmpp_sde,
            KNOWN_IMAGE_SAMPLERS.lcm,
            KNOWN_IMAGE_SAMPLERS.uni_pc,
            KNOWN_IMAGE_SAMPLERS.DDIM,
        }
    )
)

The samplers worth offering by default, settled by review rather than derived from the table.

Between them these cover the choices that actually differ: a cheap deterministic default, an ancestral one, the multistep workhorse and its SDE variant, a second-order SDE, the few-step distilled case, a predictor-corrector, and the reference solver.

SAMPLER_PRESENTATION_TIERS module-attribute 

SAMPLER_PRESENTATION_TIERS: Mapping[
    KNOWN_IMAGE_SAMPLERS, SAMPLER_PRESENTATION_TIER
] = MappingProxyType(
    {
        sampler: (presentation_tier(sampler))
        for sampler in KNOWN_IMAGE_SAMPLERS
    }
)

The presentation tier of every known sampler, covering the vocabulary exhaustively.

SAMPLER_RECOMMENDATIONS module-attribute 

SAMPLER_RECOMMENDATIONS: tuple[
    SamplerRecommendation, ...
] = (
    SamplerRecommendation(
        samplers=(),
        schedulers=(
            KNOWN_IMAGE_SCHEDULERS.karras,
            KNOWN_IMAGE_SCHEDULERS.normal,
        ),
        provenance=CONSTRAINT_PROVENANCE.upstream_author,
        source=_COMFYANONYMOUS_DISCUSSION_227,
        summary="karras and normal are the schedules to use for most samplers.",
    ),
    SamplerRecommendation(
        samplers=(KNOWN_IMAGE_SAMPLERS.DDIM,),
        schedulers=(KNOWN_IMAGE_SCHEDULERS.ddim_uniform,),
        provenance=CONSTRAINT_PROVENANCE.upstream_author,
        source=_COMFYANONYMOUS_DISCUSSION_227,
        summary="ddim_uniform is the schedule ddim is meant to be used with, matching the reference implementation.",
    ),
    SamplerRecommendation(
        samplers=(),
        schedulers=(KNOWN_IMAGE_SCHEDULERS.simple,),
        provenance=CONSTRAINT_PROVENANCE.upstream_author,
        source=_COMFYANONYMOUS_DISCUSSION_227,
        summary="simple worked well in some scenarios, such as the second pass of a hires fix.",
    ),
    SamplerRecommendation(
        samplers=(
            KNOWN_IMAGE_SAMPLERS.k_dpmpp_2m,
            KNOWN_IMAGE_SAMPLERS.dpmpp_2m_sde,
        ),
        schedulers=(KNOWN_IMAGE_SCHEDULERS.karras,),
        provenance=CONSTRAINT_PROVENANCE.community,
        source=_PROJECT_VISUAL_RULING,
        summary="The dpmpp_2m family is commonly paired with karras for smooth gradients and clean surfaces.",
    ),
    SamplerRecommendation(
        samplers=(
            KNOWN_IMAGE_SAMPLERS.k_euler,
            KNOWN_IMAGE_SAMPLERS.k_euler_a,
        ),
        schedulers=(
            KNOWN_IMAGE_SCHEDULERS.normal,
            KNOWN_IMAGE_SCHEDULERS.simple,
        ),
        provenance=CONSTRAINT_PROVENANCE.community,
        source=_PROJECT_VISUAL_RULING,
        summary="The euler family is commonly paired with normal or simple as an inexpensive general default.",
    ),
    SamplerRecommendation(
        samplers=(KNOWN_IMAGE_SAMPLERS.lcm,),
        schedulers=(
            KNOWN_IMAGE_SCHEDULERS.sgm_uniform,
            KNOWN_IMAGE_SCHEDULERS.simple,
        ),
        provenance=CONSTRAINT_PROVENANCE.community,
        source=_PROJECT_VISUAL_RULING,
        summary="Distilled and few-step solvers are commonly paired with a uniform schedule.",
    ),
    SamplerRecommendation(
        samplers=tuple(sorted(CFG_PP_SAMPLERS)),
        schedulers=(),
        provenance=CONSTRAINT_PROVENANCE.ai_horde_devs,
        source=_PROJECT_VISUAL_RULING,
        summary="CFG++ solvers expect a cfg_scale near 1.0 to 2.0; the usual range oversaturates them.",
    ),
    SamplerRecommendation(
        samplers=(),
        schedulers=(KNOWN_IMAGE_SCHEDULERS.karras,),
        provenance=CONSTRAINT_PROVENANCE.ai_horde_devs,
        source=_PROJECT_VISUAL_RULING,
        summary="karras is not the safe choice at low step counts. Reach for it for fine detail at a normal step budget, not because the budget is short.",
    ),
    SamplerRecommendation(
        samplers=(),
        schedulers=(
            KNOWN_IMAGE_SCHEDULERS.align_your_steps,
            KNOWN_IMAGE_SCHEDULERS.gits,
        ),
        provenance=CONSTRAINT_PROVENANCE.ai_horde_devs,
        source=_PROJECT_VISUAL_RULING,
        summary="align_your_steps and gits are recommended for low step counts, on the baselines they are defined for. Both exist to make a short step budget work.",
    ),
    SamplerRecommendation(
        samplers=(),
        schedulers=(
            KNOWN_IMAGE_SCHEDULERS.exponential,
            KNOWN_IMAGE_SCHEDULERS.kl_optimal,
        ),
        provenance=CONSTRAINT_PROVENANCE.ai_horde_devs,
        source=_PROJECT_VISUAL_RULING,
        summary="exponential and kl_optimal render coherently but shift colour noticeably, toward a red or magenta cast. Treat that as a distinctive look to choose deliberately rather than a fault.",
    ),
    SamplerRecommendation(
        samplers=(
            KNOWN_IMAGE_SAMPLERS.dpmpp_2m_sde,
            KNOWN_IMAGE_SAMPLERS.dpmpp_2m_sde_heun,
            KNOWN_IMAGE_SAMPLERS.dpmpp_3m_sde,
            KNOWN_IMAGE_SAMPLERS.k_dpmpp_sde,
        ),
        schedulers=(),
        provenance=CONSTRAINT_PROVENANCE.ai_horde_devs,
        source=_PROJECT_VISUAL_RULING,
        summary="The dpmpp SDE family stays usable with sampler_eta out to about 2.5.",
    ),
    SamplerRecommendation(
        samplers=(KNOWN_IMAGE_SAMPLERS.k_euler_a,),
        schedulers=(),
        provenance=CONSTRAINT_PROVENANCE.ai_horde_devs,
        source=_PROJECT_VISUAL_RULING,
        summary="k_euler_a collapses with sampler_eta above about 1.0.",
    ),
    SamplerRecommendation(
        samplers=(),
        schedulers=(KNOWN_IMAGE_SCHEDULERS.beta,),
        provenance=CONSTRAINT_PROVENANCE.ai_horde_devs,
        source=_PROJECT_VISUAL_RULING,
        summary="beta shows a colour cast and posterisation at low step counts; at a normal step budget it renders cleanly with no cast.",
    ),
    SamplerRecommendation(
        samplers=(),
        schedulers=(KNOWN_IMAGE_SCHEDULERS.ddim_uniform,),
        provenance=CONSTRAINT_PROVENANCE.ai_horde_devs,
        source=_PROJECT_VISUAL_RULING,
        summary="ddim_uniform degrades to glitchy static at very low step counts, but at a normal step budget it is not noticeably worse than simple.",
    ),
    SamplerRecommendation(
        samplers=(),
        schedulers=(
            KNOWN_IMAGE_SCHEDULERS.linear_quadratic,
        ),
        provenance=CONSTRAINT_PROVENANCE.ai_horde_devs,
        source=_PROJECT_VISUAL_RULING,
        summary="linear_quadratic blurs heavily at every tested step count on both SD1.5 and SDXL, and stays largely unusable even at a normal step budget.",
    ),
)

Advisory pairings and settings, each carrying the provenance of the claim.

Nothing here is enforced. The community entries in particular are folklore recorded for completeness, not guidance the backend's authors have endorsed. The ai_horde_devs entries were settled by looking at rendered output and are authoritative for this project.

Every schedule-character claim that went to review is recorded here at ai_horde_devs; beta, ddim_uniform and linear_quadratic initially returned unsure for want of rendered examples and were settled in a supplemental review of dedicated renders.

CONSTRAINT_PROVENANCE

Bases: StrEnum

Where a statement about a sampler came from, and therefore how much weight it carries.

Source code in horde_sdk/generation_parameters/image/constraints.py 
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
class CONSTRAINT_PROVENANCE(StrEnum):
    """Where a statement about a sampler came from, and therefore how much weight it carries."""

    upstream_author = auto()
    """Stated by the image backend's own author. The strongest non-code source."""

    community = auto()
    """Third-party folklore. Widely repeated, not endorsed by the backend's authors, unverified here."""

    measured = auto()
    """Timed on real hardware through the production render pipeline, and reproducible from its artifact.

    A measurement describes the card, resolution and pipeline it was taken on. It is evidence about
    cost, not a rule about it.
    """

    ai_horde_devs = auto()
    """Subjective opinion by the AI Horde developers."""

upstream_author class-attribute instance-attribute 

upstream_author = auto()

Stated by the image backend's own author. The strongest non-code source.

community class-attribute instance-attribute 

community = auto()

Third-party folklore. Widely repeated, not endorsed by the backend's authors, unverified here.

measured class-attribute instance-attribute 

measured = auto()

Timed on real hardware through the production render pipeline, and reproducible from its artifact.

A measurement describes the card, resolution and pipeline it was taken on. It is evidence about cost, not a rule about it.

ai_horde_devs class-attribute instance-attribute 

ai_horde_devs = auto()

Subjective opinion by the AI Horde developers.

SAMPLER_SOLVER_KNOB

Bases: StrEnum

The solver knobs a request may set, named as the backend's solver functions name them.

Source code in horde_sdk/generation_parameters/image/constraints.py 
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
class SAMPLER_SOLVER_KNOB(StrEnum):
    """The solver knobs a request may set, named as the backend's solver functions name them."""

    eta = auto()
    """Stochastic strength. At zero an SDE solver collapses onto its deterministic twin."""

    s_noise = auto()
    """Multiplier on the noise added per step."""

    s_churn = auto()
    """Total extra noise injected across the run, spread over the steps within the churn window."""

    s_tmin = auto()
    """Lower sigma bound of the churn window."""

    s_tmax = auto()
    """Upper sigma bound of the churn window."""

    solver_type = auto()
    """Which second-order correction the solver applies. The vocabulary differs per sampler."""

    order = auto()
    """Solver order: how many evaluations the step is built from, or may reuse.

    The solvers spell it differently. The single-step solvers declare `order`; the multistep ones declare
    `max_order`, which is the same concept under another name, and each record carries the spelling its
    solver uses. The sa_solver family is excluded rather than renamed: it splits the concept into a
    predictor order and a corrector order, which one number cannot express.
    """

eta class-attribute instance-attribute 

eta = auto()

Stochastic strength. At zero an SDE solver collapses onto its deterministic twin.

s_noise class-attribute instance-attribute 

s_noise = auto()

Multiplier on the noise added per step.

s_churn class-attribute instance-attribute 

s_churn = auto()

Total extra noise injected across the run, spread over the steps within the churn window.

s_tmin class-attribute instance-attribute 

s_tmin = auto()

Lower sigma bound of the churn window.

s_tmax class-attribute instance-attribute 

s_tmax = auto()

Upper sigma bound of the churn window.

solver_type class-attribute instance-attribute 

solver_type = auto()

Which second-order correction the solver applies. The vocabulary differs per sampler.

order class-attribute instance-attribute 

order = auto()

Solver order: how many evaluations the step is built from, or may reuse.

The solvers spell it differently. The single-step solvers declare order; the multistep ones declare max_order, which is the same concept under another name, and each record carries the spelling its solver uses. The sa_solver family is excluded rather than renamed: it splits the concept into a predictor order and a corrector order, which one number cannot express.

SAMPLER_PRESENTATION_TIER

Bases: StrEnum

How prominently a sampler is worth offering, now that the vocabulary is long enough to overwhelm.

This is a presentation concern and nothing else: every sampler here is accepted, priced and dispatched identically whatever its tier. A client may default to showing only the recommended tier and put the rest behind an "advanced" affordance, which is what the split is for.

Source code in horde_sdk/generation_parameters/image/constraints.py 
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
class SAMPLER_PRESENTATION_TIER(StrEnum):
    """How prominently a sampler is worth offering, now that the vocabulary is long enough to overwhelm.

    This is a presentation concern and nothing else: every sampler here is accepted, priced and
    dispatched identically whatever its tier. A client may default to showing only the recommended tier
    and put the rest behind an "advanced" affordance, which is what the split is for.
    """

    recommended = auto()
    """Worth offering by default: together these span the meaningful choices a requester has."""

    advanced = auto()
    """Legitimate, but largely an equivalent variant of something in the recommended tier.

    Nothing here is deprecated or discouraged. These are for a requester who already knows which one
    they want, and listing them all by default costs more in choice than it returns in capability.
    """

recommended class-attribute instance-attribute 

recommended = auto()

Worth offering by default: together these span the meaningful choices a requester has.

advanced class-attribute instance-attribute 

advanced = auto()

Legitimate, but largely an equivalent variant of something in the recommended tier.

Nothing here is deprecated or discouraged. These are for a requester who already knows which one they want, and listing them all by default costs more in choice than it returns in capability.

KNOWN_SAMPLER_SOLVER_TYPES

Bases: StrEnum

Every solver_type value any sampler accepts.

No sampler accepts all of them. The midpoint/heun pair and the phi_1/phi_2 pair belong to different solver families; see SamplerConstraints.solver_type_choices.

Source code in horde_sdk/generation_parameters/image/constraints.py 
121
122
123
124
125
126
127
128
129
130
131
132
class KNOWN_SAMPLER_SOLVER_TYPES(StrEnum):
    """Every `solver_type` value any sampler accepts.

    No sampler accepts all of them. The `midpoint`/`heun` pair and the `phi_1`/`phi_2` pair belong to
    different solver families; see
    [`SamplerConstraints.solver_type_choices`][horde_sdk.generation_parameters.image.constraints.SamplerConstraints.solver_type_choices].
    """

    midpoint = auto()
    heun = auto()
    phi_1 = auto()
    phi_2 = auto()

midpoint class-attribute instance-attribute 

midpoint = auto()

heun class-attribute instance-attribute 

heun = auto()

phi_1 class-attribute instance-attribute 

phi_1 = auto()

phi_2 class-attribute instance-attribute 

phi_2 = auto()

CONSTRAINT_VIOLATION_KIND

Bases: StrEnum

Why a request cannot be served as asked.

Source code in horde_sdk/generation_parameters/image/constraints.py 
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
class CONSTRAINT_VIOLATION_KIND(StrEnum):
    """Why a request cannot be served as asked."""

    knob_inapplicable = auto()
    """The sampler's solver function takes no keyword of that name, so setting it would do nothing."""

    knob_out_of_range = auto()
    """The value falls outside the range the backend declares for that sampler and knob."""

    solver_type_unsupported = auto()
    """The sampler accepts a `solver_type`, but not this one."""

    sampler_scheduler_rejected = auto()
    """The pairing is known to produce no usable image at any step count."""

    scheduler_baseline_unsupported = auto()
    """The scheduler has no definition for this baseline, so it cannot generate sigmas for it."""

knob_inapplicable class-attribute instance-attribute 

knob_inapplicable = auto()

The sampler's solver function takes no keyword of that name, so setting it would do nothing.

knob_out_of_range class-attribute instance-attribute 

knob_out_of_range = auto()

The value falls outside the range the backend declares for that sampler and knob.

solver_type_unsupported class-attribute instance-attribute 

solver_type_unsupported = auto()

The sampler accepts a solver_type, but not this one.

sampler_scheduler_rejected class-attribute instance-attribute 

sampler_scheduler_rejected = auto()

The pairing is known to produce no usable image at any step count.

scheduler_baseline_unsupported class-attribute instance-attribute 

scheduler_baseline_unsupported = auto()

The scheduler has no definition for this baseline, so it cannot generate sigmas for it.

NumericKnobRange dataclass

Represents the accepted range of one numeric solver knob for one sampler.

Attributes:

  • minimum (float) –

    Smallest accepted value, inclusive.

  • maximum (float) –

    Largest accepted value, inclusive. May be infinite.

  • default (float) –

    The value the backend uses when the knob is left unset.

  • integral (bool) –

    Whether the knob only accepts whole numbers.

  • backend_keyword (str | None) –

    The keyword the solver function itself declares, when it differs from the knob's own name. None means the two agree. This exists because the same concept is spelled differently between solvers: the multistep solvers call their order max_order, and a request that named only order would be silently dropped by the backend's own filter.

Source code in horde_sdk/generation_parameters/image/constraints.py 
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
@dataclass(frozen=True)
class NumericKnobRange:
    """Represents the accepted range of one numeric solver knob for one sampler.

    Attributes:
        minimum: Smallest accepted value, inclusive.
        maximum: Largest accepted value, inclusive. May be infinite.
        default: The value the backend uses when the knob is left unset.
        integral: Whether the knob only accepts whole numbers.
        backend_keyword: The keyword the solver function itself declares, when it differs from the knob's
            own name. `None` means the two agree. This exists because the same concept is spelled
            differently between solvers: the multistep solvers call their order `max_order`, and a
            request that named only `order` would be silently dropped by the backend's own filter.

    """

    minimum: float
    maximum: float
    default: float
    integral: bool = False
    backend_keyword: str | None = None

    def contains(self, value: float) -> bool:
        """Return whether the value is one this knob accepts."""
        if self.integral and value != int(value):
            return False

        return self.minimum <= value <= self.maximum

minimum instance-attribute 

minimum: float

maximum instance-attribute 

maximum: float

default instance-attribute 

default: float

integral class-attribute instance-attribute 

integral: bool = False

backend_keyword class-attribute instance-attribute 

backend_keyword: str | None = None

__init__ 

__init__(
    minimum: float,
    maximum: float,
    default: float,
    integral: bool = False,
    backend_keyword: str | None = None,
) -> None

contains 

contains(value: float) -> bool

Return whether the value is one this knob accepts.

Source code in horde_sdk/generation_parameters/image/constraints.py 
176
177
178
179
180
181
def contains(self, value: float) -> bool:
    """Return whether the value is one this knob accepts."""
    if self.integral and value != int(value):
        return False

    return self.minimum <= value <= self.maximum

SamplerConstraints dataclass

Represents everything known about one sampler's knobs, cost and identity.

Attributes:

Source code in horde_sdk/generation_parameters/image/constraints.py 
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
@dataclass(frozen=True)
class SamplerConstraints:
    """Represents everything known about one sampler's knobs, cost and identity.

    Attributes:
        sampler: The sampler these constraints describe.
        backend_solver_function: The backend function the sampler resolves to, for traceability.
        numeric_knob_ranges: The numeric knobs the sampler accepts, and their ranges. A knob absent
            from this mapping is one the sampler's solver function does not take.
        solver_type_choices: The `solver_type` values the sampler accepts, empty when it takes none.
        work_profile: The sampler's marginal work shape. It is deliberately distinct from trajectory
            length and does not claim exact wall-clock cost.
        measured_cost_ratio_sd15: Measured per-step wall-clock cost relative to `k_euler` on a
            stable_diffusion_1 model at 512x512, or `None` where a ratio is meaningless. Carries a small
            positive bias on the one-evaluation samplers, because per-step host work is a visible
            fraction of a 512x512 step.
        measured_cost_ratio_sdxl: The same ratio on a stable_diffusion_xl model at 1024x1024. The larger
            step swamps the per-step host work, so this is the figure that reflects what the sampler
            itself costs. See
            [`MEASURED_COST_RATIO_PROVENANCE`][horde_sdk.generation_parameters.image.constraints.MEASURED_COST_RATIO_PROVENANCE].

    """

    sampler: KNOWN_IMAGE_SAMPLERS
    backend_solver_function: str
    numeric_knob_ranges: Mapping[SAMPLER_SOLVER_KNOB, NumericKnobRange]
    solver_type_choices: tuple[KNOWN_SAMPLER_SOLVER_TYPES, ...]
    work_profile: SamplerWorkProfile
    measured_cost_ratio_sd15: float | None
    measured_cost_ratio_sdxl: float | None

    def accepts_knob(self, knob: SAMPLER_SOLVER_KNOB) -> bool:
        """Return whether the sampler's solver function takes this knob."""
        if knob is SAMPLER_SOLVER_KNOB.solver_type:
            return bool(self.solver_type_choices)

        return knob in self.numeric_knob_ranges

sampler instance-attribute 

sampler: KNOWN_IMAGE_SAMPLERS

backend_solver_function instance-attribute 

backend_solver_function: str

numeric_knob_ranges instance-attribute 

numeric_knob_ranges: Mapping[
    SAMPLER_SOLVER_KNOB, NumericKnobRange
]

solver_type_choices instance-attribute 

solver_type_choices: tuple[KNOWN_SAMPLER_SOLVER_TYPES, ...]

work_profile instance-attribute 

work_profile: SamplerWorkProfile

measured_cost_ratio_sd15 instance-attribute 

measured_cost_ratio_sd15: float | None

measured_cost_ratio_sdxl instance-attribute 

measured_cost_ratio_sdxl: float | None

__init__ 

__init__(
    sampler: KNOWN_IMAGE_SAMPLERS,
    backend_solver_function: str,
    numeric_knob_ranges: Mapping[
        SAMPLER_SOLVER_KNOB, NumericKnobRange
    ],
    solver_type_choices: tuple[
        KNOWN_SAMPLER_SOLVER_TYPES, ...
    ],
    work_profile: SamplerWorkProfile,
    measured_cost_ratio_sd15: float | None,
    measured_cost_ratio_sdxl: float | None,
) -> None

accepts_knob 

accepts_knob(knob: SAMPLER_SOLVER_KNOB) -> bool

Return whether the sampler's solver function takes this knob.

Source code in horde_sdk/generation_parameters/image/constraints.py 
215
216
217
218
219
220
def accepts_knob(self, knob: SAMPLER_SOLVER_KNOB) -> bool:
    """Return whether the sampler's solver function takes this knob."""
    if knob is SAMPLER_SOLVER_KNOB.solver_type:
        return bool(self.solver_type_choices)

    return knob in self.numeric_knob_ranges

ConstraintViolation dataclass

Represents one reason a request cannot be served as asked.

Attributes:

Source code in horde_sdk/generation_parameters/image/constraints.py 
223
224
225
226
227
228
229
230
231
232
233
234
@dataclass(frozen=True)
class ConstraintViolation:
    """Represents one reason a request cannot be served as asked.

    Attributes:
        kind: Which rule the request breaks.
        detail: A sentence naming the offending value and what would be accepted instead.

    """

    kind: CONSTRAINT_VIOLATION_KIND
    detail: str

kind instance-attribute 

kind: CONSTRAINT_VIOLATION_KIND

detail instance-attribute 

detail: str

__init__ 

__init__(
    kind: CONSTRAINT_VIOLATION_KIND, detail: str
) -> None

SamplerRecommendation dataclass

Represents one advisory statement about how a sampler is best used.

Recommendations never block a request. They exist so clients can offer a sensible default and can say where the suggestion came from.

Attributes:

Source code in horde_sdk/generation_parameters/image/constraints.py 
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
@dataclass(frozen=True)
class SamplerRecommendation:
    """Represents one advisory statement about how a sampler is best used.

    Recommendations never block a request. They exist so clients can offer a sensible default and can
    say where the suggestion came from.

    Attributes:
        samplers: The samplers the statement applies to, empty when it applies to all of them.
        schedulers: The schedulers the statement recommends, empty when it recommends none.
        provenance: Whose statement this is.
        source: A human-readable citation.
        summary: The statement itself.

    """

    samplers: tuple[KNOWN_IMAGE_SAMPLERS, ...]
    schedulers: tuple[KNOWN_IMAGE_SCHEDULERS, ...]
    provenance: CONSTRAINT_PROVENANCE
    source: str
    summary: str

samplers instance-attribute 

samplers: tuple[KNOWN_IMAGE_SAMPLERS, ...]

schedulers instance-attribute 

schedulers: tuple[KNOWN_IMAGE_SCHEDULERS, ...]

provenance instance-attribute 

provenance: CONSTRAINT_PROVENANCE

source instance-attribute 

source: str

summary instance-attribute 

summary: str

__init__ 

__init__(
    samplers: tuple[KNOWN_IMAGE_SAMPLERS, ...],
    schedulers: tuple[KNOWN_IMAGE_SCHEDULERS, ...],
    provenance: CONSTRAINT_PROVENANCE,
    source: str,
    summary: str,
) -> None

presentation_tier 

presentation_tier(
    sampler: KNOWN_IMAGE_SAMPLERS,
) -> SAMPLER_PRESENTATION_TIER

Return how prominently a sampler is worth offering.

Parameters:

Returns:

Source code in horde_sdk/generation_parameters/image/constraints.py 
758
759
760
761
762
763
764
765
766
767
768
769
770
771
def presentation_tier(sampler: KNOWN_IMAGE_SAMPLERS) -> SAMPLER_PRESENTATION_TIER:
    """Return how prominently a sampler is worth offering.

    Args:
        sampler: The sampler to look up.

    Returns:
        `recommended` for the default-offer set, `advanced` for everything else.

    """
    if sampler in RECOMMENDED_SAMPLERS:
        return SAMPLER_PRESENTATION_TIER.recommended

    return SAMPLER_PRESENTATION_TIER.advanced

get_sampler_constraints 

get_sampler_constraints(
    sampler: KNOWN_IMAGE_SAMPLERS,
) -> SamplerConstraints

Return the constraint record for a sampler.

Parameters:

Returns:

Raises:

  • KeyError –

    If the sampler has no record, which means the table has drifted from the enum.

Source code in horde_sdk/generation_parameters/image/constraints.py 
928
929
930
931
932
933
934
935
936
937
938
939
940
941
def get_sampler_constraints(sampler: KNOWN_IMAGE_SAMPLERS) -> SamplerConstraints:
    """Return the constraint record for a sampler.

    Args:
        sampler: The sampler to look up.

    Returns:
        The record describing that sampler's knobs, cost and identity.

    Raises:
        KeyError: If the sampler has no record, which means the table has drifted from the enum.

    """
    return SAMPLER_CONSTRAINTS[sampler]

is_knob_applicable 

is_knob_applicable(
    sampler: KNOWN_IMAGE_SAMPLERS, knob: SAMPLER_SOLVER_KNOB
) -> bool

Return whether setting this knob on this sampler would have any effect.

Parameters:

Returns:

  • bool –

    Whether the sampler's solver function accepts a keyword of that name.

Source code in horde_sdk/generation_parameters/image/constraints.py 
944
945
946
947
948
949
950
951
952
953
954
955
def is_knob_applicable(sampler: KNOWN_IMAGE_SAMPLERS, knob: SAMPLER_SOLVER_KNOB) -> bool:
    """Return whether setting this knob on this sampler would have any effect.

    Args:
        sampler: The sampler the request names.
        knob: The knob the request wants to set.

    Returns:
        Whether the sampler's solver function accepts a keyword of that name.

    """
    return SAMPLER_CONSTRAINTS[sampler].accepts_knob(knob)

applicable_knobs 

applicable_knobs(
    sampler: KNOWN_IMAGE_SAMPLERS,
) -> frozenset[SAMPLER_SOLVER_KNOB]

Return every knob this sampler accepts.

Parameters:

Returns:

Source code in horde_sdk/generation_parameters/image/constraints.py 
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
def applicable_knobs(sampler: KNOWN_IMAGE_SAMPLERS) -> frozenset[SAMPLER_SOLVER_KNOB]:
    """Return every knob this sampler accepts.

    Args:
        sampler: The sampler to look up.

    Returns:
        The knobs a request may set for it.

    """
    constraints = SAMPLER_CONSTRAINTS[sampler]
    knobs = set(constraints.numeric_knob_ranges)

    if constraints.solver_type_choices:
        knobs.add(SAMPLER_SOLVER_KNOB.solver_type)

    return frozenset(knobs)

is_scheduler_applicable 

is_scheduler_applicable(
    scheduler: KNOWN_IMAGE_SCHEDULERS,
    baseline: KNOWN_IMAGE_GENERATION_BASELINE,
) -> bool

Return whether a scheduler can generate sigmas for a baseline.

Parameters:

  • scheduler (KNOWN_IMAGE_SCHEDULERS) –

    The scheduler the request names.

  • baseline (KNOWN_IMAGE_GENERATION_BASELINE) –

    The baseline of the model the request names.

Returns:

  • bool –

    Whether the pairing is defined. Schedulers with no baseline restriction are always applicable.

Source code in horde_sdk/generation_parameters/image/constraints.py 
977
978
979
980
981
982
983
984
985
986
987
988
989
990
def is_scheduler_applicable(scheduler: KNOWN_IMAGE_SCHEDULERS, baseline: KNOWN_IMAGE_GENERATION_BASELINE) -> bool:
    """Return whether a scheduler can generate sigmas for a baseline.

    Args:
        scheduler: The scheduler the request names.
        baseline: The baseline of the model the request names.

    Returns:
        Whether the pairing is defined. Schedulers with no baseline restriction are always applicable.

    """
    allowed_baselines = SCHEDULER_BASELINE_APPLICABILITY.get(scheduler)

    return allowed_baselines is None or baseline in allowed_baselines

list_constraint_violations 

list_constraint_violations(
    *,
    sampler: KNOWN_IMAGE_SAMPLERS,
    scheduler: KNOWN_IMAGE_SCHEDULERS | None = None,
    baseline: KNOWN_IMAGE_GENERATION_BASELINE | None = None,
    numeric_knobs: (
        Mapping[SAMPLER_SOLVER_KNOB, float] | None
    ) = None,
    solver_type: KNOWN_SAMPLER_SOLVER_TYPES | None = None
) -> list[ConstraintViolation]

Return every reason a request cannot be served exactly as asked.

An empty list means the combination is renderable. A non-empty one is grounds for rejection rather than substitution: the caller asked for something the backend cannot do, and quietly doing something else would return an image nobody requested.

Parameters:

  • sampler (KNOWN_IMAGE_SAMPLERS) –

    The sampler the request names.

  • scheduler (KNOWN_IMAGE_SCHEDULERS | None, default: None ) –

    The scheduler the request resolved to, if any.

  • baseline (KNOWN_IMAGE_GENERATION_BASELINE | None, default: None ) –

    The baseline of the model the request names, if known.

  • numeric_knobs (Mapping[SAMPLER_SOLVER_KNOB, float] | None, default: None ) –

    The numeric solver knobs the request set, keyed by knob.

  • solver_type (KNOWN_SAMPLER_SOLVER_TYPES | None, default: None ) –

    The solver type the request set, if any.

Returns:

Raises:

  • KeyError –

    If the sampler has no constraint record.

Source code in horde_sdk/generation_parameters/image/constraints.py 
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
def list_constraint_violations(
    *,
    sampler: KNOWN_IMAGE_SAMPLERS,
    scheduler: KNOWN_IMAGE_SCHEDULERS | None = None,
    baseline: KNOWN_IMAGE_GENERATION_BASELINE | None = None,
    numeric_knobs: Mapping[SAMPLER_SOLVER_KNOB, float] | None = None,
    solver_type: KNOWN_SAMPLER_SOLVER_TYPES | None = None,
) -> list[ConstraintViolation]:
    """Return every reason a request cannot be served exactly as asked.

    An empty list means the combination is renderable. A non-empty one is grounds for rejection rather
    than substitution: the caller asked for something the backend cannot do, and quietly doing something
    else would return an image nobody requested.

    Args:
        sampler: The sampler the request names.
        scheduler: The scheduler the request resolved to, if any.
        baseline: The baseline of the model the request names, if known.
        numeric_knobs: The numeric solver knobs the request set, keyed by knob.
        solver_type: The solver type the request set, if any.

    Returns:
        One entry per broken rule, in the order the rules are checked.

    Raises:
        KeyError: If the sampler has no constraint record.

    """
    constraints = SAMPLER_CONSTRAINTS[sampler]
    violations: list[ConstraintViolation] = []

    if numeric_knobs:
        violations.extend(_numeric_knob_violations(constraints, numeric_knobs))

    if solver_type is not None:
        violations.extend(_solver_type_violations(constraints, solver_type))

    if scheduler is not None and (sampler, scheduler) in REJECTED_SAMPLER_SCHEDULER_PAIRINGS:
        violations.append(
            ConstraintViolation(
                kind=CONSTRAINT_VIOLATION_KIND.sampler_scheduler_rejected,
                detail=f"{sampler} does not converge on the {scheduler} schedule.",
            ),
        )

    if scheduler is not None and baseline is not None and not is_scheduler_applicable(scheduler, baseline):
        allowed = ", ".join(sorted(SCHEDULER_BASELINE_APPLICABILITY[scheduler]))
        violations.append(
            ConstraintViolation(
                kind=CONSTRAINT_VIOLATION_KIND.scheduler_baseline_unsupported,
                detail=f"{scheduler} is only defined for {allowed}, not {baseline}.",
            ),
        )

    return violations