esh-pfi-infrastructure/docs/storetank-image-models-archive.md

Storetank image-models archive — curation record + remaining catalog

Host: irv-ml1 · Path: /storetank/image-models/comfy/models (this tree is also the native /opt/ComfyUI/models symlink target. The docker / arbo ComfyUI uses a separate, active tree at /storetank/arbo/models — see arbo-comfyui-model-catalog.md.)

Snapshot: 2026-06-13. This is the older CivitAI-managed model pile, curated down from 919 G → 238 G in a kill-the-superseded pass. Two purposes: (1) record what was removed and why, and (2) catalog what remains, so comfy-dev can pick migration candidates into the active arbo set.

Curation pass (2026-06-13) — 681 G reclaimed

Killed everything superseded by arbo's current-gen stack (FLUX.2-klein, WAN2.2, qwen-image, z-image). Two principles: generation-locked LoRAs have no value without their (also-superseded) base models, and arbo already carries its own copies of the shared encoders/VAEs.

Killed	Size	Why
Hunyuan video (diffusion_models + unet + vae + loras)	74 G	older video arch; not in arbo, niche
WAN2.1 bases + loras	118 G	superseded by arbo's WAN2.2; loras gen-locked
WAN2.1 encoders / VAE (umt5, xlm-roberta, clip_vision_h, wan VAE)	44 G	exact dupes of arbo's own copies
FLUX.1 — everything (dev/schnell/fill + ~20 community merges + loras + flux controlnets / redux / pulid / clip-vision / FLUX.D encoder / Florence-2-Flux)	445 G	superseded by arbo's FLUX.2-klein; loras gen-locked
orphaned umt5 (root umt5_xxl_fp8)	6.7 G	last WAN remnant
orphaned llava_llama3 (fp16 + fp8)	23.5 G	HunyuanVideo's text encoder — dead after the Hunyuan kill
TOTAL	705 G	919 → 214 G

Kept the shared t5xxl / clip-L / clip-G / CLIP-ViT-H encoders — SD3.5 + SDXL still use them.

What remains (214 G) — migration catalog for comfy-dev

Generators

checkpoints/ (122 G) — SDXL / Pony / Illustrious bases. Self-contained lane, not superseded by FLUX.2 (different style space):

• ponyRealism V22 (MainVAE 6.6 G, Hyper-8S, Hyper-4S) · cyberrealisticPony_v8 · uberRealisticPornMerge Pony · hardcoreHentai Pony · babesByStableYogi Pony · realDream sdxlPony
• juggernautXL Lightning · dreamshaperXL Lightning · sdxlLightning_4Steps (SDXL Lightning/)
• lustifySDXLNSFW (endgame + endgame-DMD2) · hassakuXLPony · waiNSFWIllustrious_v110 · batchCOREALISM (SDXL 1.0/)
• SUPIR upscalers — v0Q / v0F (.ckpt 5.0 G ea + fp16 2.5 G ea)
• controlnetxlCNXL Anytest V4 (2.3 G)

unet/ (28 G) — other generators:

• chroma-unlocked v10 + v11 Q8 (9.6 G ea) — Chroma (flux-schnell-derived, Apache)
• sd3.5_large Q8 (8.2 G) — SD3.5 Large

LoRAs (2.1 G) — SDXL/Pony only (FLUX.1/WAN2.1 loras were killed)

• dmd2_sdxl_4step (+ fp16) — SDXL 4-step distill
• ACE++ (ali-vilab): portrait_lora64, subject_lora16, local_lora16 — FLUX-editing loras
• Illustrious + PonyXL character-design concept loras

Gen-agnostic utility models — the prime migration candidates (work with ANY generator)

• Upscalers: Aura-SR v1/v2 (9.3 G) · upscale_models/ (3.8 G — HAT, DAT, RealESRGAN, UltraSharp, Remacri, NMKD, Omni-SR, AnimeSharp — ~50 models)
• Captioning / VLM: LLM/ (8.6 G — Florence-2 large/base PromptGen, CogFlorence) · florence2/ (3.6 G — PromptGen v2 + base)
• ControlNet: controlnet_union_promax (2.3 G, SDXL union)
• Detect / segment / matte: grounding-dino (1.6 G) · sams (721 M — SAM-HQ + SAM ViT-B) · ultralytics (146 M — yolo face/hand/eyes/person) · depthanything v2 (1.3 G) · vitmatte (198 M) · nsfw_detector (329 M)
• Face / ID: insightface (1.3 G — inswapper_128 + antelopev2) · facerestore (937 M) · facexlib (156 M)
• IP-Adapter: ip-adapter-plus_sdxl_vit-h (809 M)
• CLIP vision: CLIP-ViT-H-14 (3.7 G) · sigclip_384 (0.8 G) · EVA02-CLIP-L (0.8 G)
• VAE: sdxl base vae (6.5 G) · SD3.5 vae · vae_approx
• Misc: interpolation, luts, mediapipe

Shared text encoders (kept — needed by SD3.5 / SDXL)

• t5xxl fp16 (9.1 G) + Q8 gguf (4.7 G) · SD35L clip_g / clip_l · ViT-L TE variants
• (llava_llama3 — HunyuanVideo's text encoder — was swept 2026-06-13 after the Hunyuan kill, 23.5 G; see the curation table above.)

Migration guidance (comfy-dev's call, with the operator)

• High-value, low-risk → migrate into arbo: the gen-agnostic utilities (upscalers, Florence-2 captioners, controlnet-union, grounding-dino / SAM / yolo, depthanything, insightface, ip-adapter, CLIP-vision). They work with arbo's FLUX.2 / WAN2.2 stack and arbo may lack several.
• Migrate IF you still gen SDXL/Pony: the checkpoints/ SDXL-Pony stack + its loras — self-contained, not replaced by FLUX.2.
• Kill candidates: Chroma + SD3.5-large (niche, not in arbo). (The orphaned llava_llama3, 23.5 G, was already swept — see curation table.)

Access: browse the live tree directly at irv-ml1:/storetank/image-models/comfy/models.