mirror of
https://github.com/Comfy-Org/ComfyUI_frontend.git
synced 2026-06-10 16:29:26 +00:00
87625d852b
Shared data layer for the Model Library feature, with no app-visible UI: - Asset metadata/sort/filter helpers (assetService, assetMetadataUtils, assetSortUtils, filterTypes) extended for model browsing. - Base-model inference with curated overrides (baseModelInference, baseModelOverrides, baseModelCategoryOverrides, comfyOrgProviderOverrides). - Model grouping and provider formatting (modelGroups), category icon helpers (categoryUtil), and category placeholder support. - Model Library data sources: cloud (useModelLibrarySource) and on-disk local (useLocalModelLibrarySource), plus useRecentlyUsedModels and the node-preview drag image helper. Consumers land in stacked follow-up PRs (sidebar tab, in-node picker). Until then their orphaned files/exports are scoped under the existing knip stacked-PR conventions (ignore list + @knipIgnoreUsedByStackedPR), to be removed as each consumer arrives.
211 lines
8.3 KiB
TypeScript
211 lines
8.3 KiB
TypeScript
/**
|
|
* Filename-based base-model inference for assets that lack both a
|
|
* `metadata.base_model` field and a [[BASE_MODEL_OVERRIDES]] entry — typically
|
|
* Civitai-sourced LoRAs with no HuggingFace repo. The pattern set mirrors the
|
|
* Python scraper's canonical rules so a `flux1-…` LoRA, a `zimage-…` LoRA, etc.
|
|
* land in the right bucket without manual tagging.
|
|
*
|
|
* Underscores are normalised to hyphens before matching because `\b` treats
|
|
* `_` as a word char and would otherwise miss `qwen-image_lora`.
|
|
*/
|
|
|
|
const CANONICAL_RULES: ReadonlyArray<
|
|
readonly [label: string, pattern: RegExp]
|
|
> = [
|
|
// Flux family — longest match first
|
|
['Flux.2 Klein', /\bflux[-.\s]?2[-.\s]?klein\b/i],
|
|
['Flux.2 dev', /\bflux[-.\s]?2\b/i],
|
|
['Flux.1 Krea', /\bflux[-.\s]?1?[-.\s]?krea\b/i],
|
|
['Flux.1 Kontext', /\bflux[-.\s]?1?[-.\s]?kontext\b/i],
|
|
['Flux.1 Redux', /\bflux[-.\s]?1?[-.\s]?redux\b/i],
|
|
['Flux.1 Schnell', /\bflux[-.\s]?1?[-.\s]?schnell\b/i],
|
|
['Flux.1 dev', /\bflux[-.\s]?1\b/i],
|
|
['Flux.1 dev', /\bflux\b/i],
|
|
// Stable Diffusion family — require sd/stable_diffusion prefix
|
|
['SDXL', /\bsd[-.\s]?xl\b|\bstable[-.\s]?diffusion[-.\s]?xl\b/i],
|
|
['SD 3.5', /\b(?:sd|stable[-.\s]?diffusion)[-.\s]?v?3[-.\s]?\.?5\b/i],
|
|
['SD 3', /\b(?:sd|stable[-.\s]?diffusion)[-.\s]?v?3\b/i],
|
|
['SD 2.1', /\b(?:sd|stable[-.\s]?diffusion)[-.\s]?v?2[-.\s]?\.?1\b/i],
|
|
['SD 2', /\b(?:sd|stable[-.\s]?diffusion)[-.\s]?v?2\b/i],
|
|
['SD 1.5', /\b(?:sd|stable[-.\s]?diffusion)[-.\s]?v?1[-.\s]?\.?5\b/i],
|
|
// Wan
|
|
['Wan 2.2', /\bwan[-.\s]?2[-.\s]?\.?2\b/i],
|
|
['Wan 2.1', /\bwan[-.\s]?2[-.\s]?\.?1\b/i],
|
|
['Wan', /\bwan\b/i],
|
|
// Hunyuan
|
|
['HunyuanVideo 1.5', /\bhunyuan[-.\s]?video[-.\s]?1[-.\s]?\.?5\b/i],
|
|
['HunyuanVideo', /\bhunyuan[-.\s]?video\b/i],
|
|
['Hunyuan Image', /\bhunyuan[-.\s]?image\b/i],
|
|
['Hunyuan 3D', /\bhunyuan[-.\s]?3d\b/i],
|
|
// Qwen — Image/Edit before plain Qwen
|
|
['Qwen Image Edit', /\bqwen[-.\s]?image[-.\s]?edit\b/i],
|
|
['Qwen Image', /\bqwen[-.\s]?image\b/i],
|
|
['Qwen', /\bqwen\b/i],
|
|
// SDXL-derivative bases — community treats as their own family
|
|
['Pony', /\bpony\b/i],
|
|
['Illustrious', /\billustrious\b/i],
|
|
// Other diffusion families — variants before family root
|
|
['HiDream I1', /\bhi[-_.\s]?dream[-_.\s]?i1\b/i],
|
|
['HiDream O1', /\bhi[-_.\s]?dream[-_.\s]?o1\b/i],
|
|
['HiDream', /\bhi[-.\s]?dream\b/i],
|
|
['Chroma1 Radiance', /\bchroma\d*[-_.\s]?radiance\b/i],
|
|
['Chroma1 HD', /\bchroma\d*[-_.\s]?hd\b/i],
|
|
['Chroma', /\bchroma\d*\b/i],
|
|
// Captioner / VLM families — placed before LTX so LTXV-packaged
|
|
// captioner files (e.g. `ltxv_florence2_promptgen_…`) classify by their
|
|
// actual model family, not the packaging prefix.
|
|
['CogFlorence', /\bcog[-_.\s]?florence\b/i],
|
|
['Florence-2', /\bflorence[-_.\s]?2\b/i],
|
|
['JoyCaption', /\bjoy[-_.\s]?caption\d*\b/i],
|
|
['LLaVA', /\bllava\b/i],
|
|
['SmolVLM', /\bsmol[-_.\s]?vlm\b/i],
|
|
['SmolLM2', /\bsmol[-_.\s]?lm\d*\b/i],
|
|
['SuperPrompt', /\bsuper[-_.\s]?prompt\b/i],
|
|
// Voice / TTS — Chatterbox Turbo before bare Chatterbox
|
|
['Chatterbox Turbo', /\bchatterbox[-_.\s]?turbo\b/i],
|
|
['Chatterbox', /\bchatterbox\b/i],
|
|
// Depth — V2 before V1
|
|
['Depth Anything V2', /\bdepth[-_.\s]?anything[-_.\s]?v?2\b/i],
|
|
['Depth Anything', /\bdepth[-_.\s]?anything\b/i],
|
|
// Other utility / motion / upscale families
|
|
['SegFormer', /\bsegformer\b/i],
|
|
['LivePortrait', /\blive[-_.\s]?portrait\b/i],
|
|
['DynamiCrafter', /\bdynami[-_.\s]?crafter\b/i],
|
|
['SeedVR2', /\bseed[-_.\s]?vr\d*\b/i],
|
|
['FlashVSR', /\bflash[-_.\s]?vsr\b/i],
|
|
['MimicMotion', /\bmimic[-_.\s]?motion\b/i],
|
|
['LatentSync', /\blatent[-_.\s]?sync\b/i],
|
|
// Vision encoders — SigLIP before CLIP so CLIP-only matches don't swallow siglip-*
|
|
['SigLIP', /\bsiglip\b/i],
|
|
['CLIP-ViT', /\bclip[-_.\s]?vit\b/i],
|
|
['Llama 3.2', /\bllama[-_.\s]?3[-_.\s]?\.?2\b/i],
|
|
['LTX 2.3', /\bltx[-.\s]?v?2[-.\s]?\.?3\b/i],
|
|
['LTX 2', /\bltx[-.\s]?v?2\b/i],
|
|
['LTX Video', /\bltx\b/i],
|
|
// Upscalers / restoration
|
|
['UltraSharp', /\bultrasharp\b/i],
|
|
['Real-ESRGAN', /\breal[-_.\s]?esrgan\b/i],
|
|
// Depth / normal estimation
|
|
['Lotus', /\blotus\b/i],
|
|
// Matting / background
|
|
['ViTMatte', /\bvit[-_.\s]?matte\b/i],
|
|
['LayerDiffusion', /\blayer[-_.\s]?diffusion\b|\blayer[-_.\s]?xl\b/i],
|
|
// Motion / interpolation
|
|
['RIFE', /\brife\b/i],
|
|
// Detection / pose
|
|
['GroundingDINO', /\bgrounding[-_.\s]?dino\b/i],
|
|
['DWPose', /\bdwpose\b|\bdw[-_.\s]?ll[-_.\s]?ucoco\b/i],
|
|
['Face Parsing', /\bface[-_.\s]?parsing\b/i],
|
|
// Additional language models
|
|
['ChatGLM3', /\bchat[-_.\s]?glm\d*\b/i],
|
|
['Gemma', /\bgemma\d*\b/i],
|
|
['Cosmos', /\bcosmos\b/i],
|
|
['Mochi', /\bmochi\b/i],
|
|
['Stable Audio', /\bstable[-.\s]?audio\b/i],
|
|
['AuraFlow', /\bauraflow\b/i],
|
|
['PixArt', /\bpixart\b/i],
|
|
['Kandinsky', /\bkandinsky\b/i],
|
|
['Playground', /\bplayground\b/i],
|
|
['Kolors', /\bkolors\b/i],
|
|
['Z-Image', /\bz[-_.\s]?image(?:[-_.\s]?turbo)?\b/i],
|
|
['Lumina', /\blumina\b/i],
|
|
['CogVideo', /\bcogvideo\b/i],
|
|
['AnimateDiff', /\banimatediff\b/i],
|
|
['ERNIE', /\bernie\b/i],
|
|
['Omnigen', /\bomnigen\d*\b/i],
|
|
['Ovis', /\bovis\b/i],
|
|
['ACE-Step', /\bace[-.\s]?step\b/i],
|
|
['HuMo', /\bhumo\b/i],
|
|
['LongCat', /\blongcat\b/i],
|
|
['Trellis', /\btrellis\b/i],
|
|
['USO', /\buso\b/i],
|
|
['OneReward', /\bone[-.\s]?reward\b/i],
|
|
['MoGe', /\bmoge\b/i],
|
|
['UltraShape', /\bultrashape\b/i],
|
|
['NewBie', /\bnewbie\b/i],
|
|
['PixelDiT', /\bpixel[-.\s]?dit\b/i],
|
|
['SAM 3D', /\bsam[-.\s]?3d\b/i],
|
|
['SAM 3', /\bsam[-.\s]?3(?!d)\b/i],
|
|
['SAM 2', /\bsam[-.\s]?2\b/i],
|
|
['SAM', /\bsam\b/i],
|
|
['BiRefNet', /\bbirefnet\b/i]
|
|
] as const
|
|
|
|
export function inferBaseModelFromText(text: string): string | null {
|
|
if (!text) return null
|
|
// Underscores are word chars to regex \b — swap to hyphens so things like
|
|
// "Qwen-Image_ComfyUI" or "flux1-foo" match cleanly.
|
|
const normalized = text.replace(/_/g, '-')
|
|
for (const [label, pattern] of CANONICAL_RULES) {
|
|
if (pattern.test(normalized)) return label
|
|
}
|
|
return null
|
|
}
|
|
|
|
const CANONICAL_LABELS: ReadonlySet<string> = new Set(
|
|
CANONICAL_RULES.map(([label]) => label)
|
|
)
|
|
|
|
/**
|
|
* Family-prefix rules. Maps labels (canonical and common non-canonical
|
|
* variants like `LTXV2`) onto a family bucket so refinement can spot when a
|
|
* filename suggests a more specific variant of the same family.
|
|
*/
|
|
const FAMILY_PREFIX_RULES: ReadonlyArray<readonly [RegExp, string]> = [
|
|
[/^(?:ltxv|ltx)/i, 'ltx'],
|
|
[/^(?:sdxl|sd|stable[-.\s]?diffusion)/i, 'sd'],
|
|
[/^flux/i, 'flux'],
|
|
[/^wan/i, 'wan'],
|
|
[/^hunyuan/i, 'hunyuan'],
|
|
[/^qwen/i, 'qwen'],
|
|
[/^z[-_.\s]?image/i, 'zimage'],
|
|
[/^hi[-_.\s]?dream/i, 'hidream'],
|
|
[/^sam/i, 'sam']
|
|
] as const
|
|
|
|
function familyOf(label: string): string {
|
|
for (const [pattern, family] of FAMILY_PREFIX_RULES) {
|
|
if (pattern.test(label)) return family
|
|
}
|
|
return label.toLowerCase().match(/^[a-z]+/)?.[0] ?? label.toLowerCase()
|
|
}
|
|
|
|
/**
|
|
* Refines metadata-derived base-model labels using filename inference. When
|
|
* the filename suggests a more specific variant of the same family — e.g.
|
|
* `LTX_2.3_…` whose HuggingFace card says only `Lightricks/LTX-Video` —
|
|
* promote to the specific variant.
|
|
*
|
|
* Rules per existing label:
|
|
* 1. If a filename-inferred label shares its family AND the existing label
|
|
* is non-canonical, replace with the canonical inferred label.
|
|
* 2. If both are canonical and same family, prefer the one with a version
|
|
* digit when the other has none.
|
|
*/
|
|
export function refineBaseModelLabels(
|
|
labels: readonly string[],
|
|
filenameSources: readonly string[]
|
|
): string[] {
|
|
if (labels.length === 0) return [...labels]
|
|
const inferences = filenameSources
|
|
.map((s) => inferBaseModelFromText(s))
|
|
.filter((x): x is string => Boolean(x))
|
|
if (inferences.length === 0) return [...labels]
|
|
return labels.map((existing) => {
|
|
const family = familyOf(existing)
|
|
for (const inferred of inferences) {
|
|
if (familyOf(inferred) !== family) continue
|
|
if (inferred === existing) return existing
|
|
const existingCanonical = CANONICAL_LABELS.has(existing)
|
|
const inferredCanonical = CANONICAL_LABELS.has(inferred)
|
|
if (!existingCanonical && inferredCanonical) return inferred
|
|
if (existingCanonical && inferredCanonical) {
|
|
const inferredHasDigit = /\d/.test(inferred)
|
|
const existingHasDigit = /\d/.test(existing)
|
|
if (inferredHasDigit && !existingHasDigit) return inferred
|
|
}
|
|
}
|
|
return existing
|
|
})
|
|
}
|