Eve library

Isis toolbox audit: environments, costumes, movement, prompting (2026-09-14)

Guide

Revision 1

On this page

Scope: what the Isis arsenal (tracker ISIS_CHROMA_RUNPOD_MVP_TODOS_2026-09-11.md, sections C, V, A, L) already covers, a verdict on each of the 27 Civitai links the user supplied, a sweep of the most-downloaded Civitai LoRAs and checkpoints on the bases Isis runs, and a Hugging Face and ComfyUI release-note pass on dance, music video, camera, wardrobe and set tools. Content boundary (user, 2026-09-14): no porn or sexual acts; music video, dance, nude yoga and classy artistic nudity are in scope. Nothing here is pulled or added to the tracker; every addition waits for the user's pick.

Method, so it can be repeated: Civitai model API with the project key (/api/v1/models/<id>, and /api/v1/models?baseModels=<base>&types=LORA|Checkpoint&sort=Most Downloaded|Highest Rated for Chroma, ZImageTurbo, ZImageBase, Wan Video 2.2 I2V-A14B and T2V-A14B, Qwen, Flux.2 Klein 4B, HiDream and LTXV 2.5: 1,166 unique models, 270 dropped as sex-act by name and tags); a version's paidAccess field marks a paid download (the 1-byte download probe answers 403); Hugging Face model API with ?blobs=true for licences and byte sizes; the ComfyUI release notes 0.25–0.35; the worker's object-info snapshot infra/runpod/endpoints/node-classes/worker-comfyui-5.10.0-video.json for node classes that already ship; four of the linked workflow files downloaded and their node graphs read.

1. What the arsenal covers today#

Area Covered by Gap
People, bodies, faces Chroma (4 checkpoints, 6 LoRAs), Z-Image base plus Turbo AIO (5 LoRAs), HiDream-O1 personalize, klein 4B and Qwen edit No face or hand detailer pass; Chroma has no ControlNet (C.12.03)
Poses Z-Image Fun union pose control, SCAIL-2 and Animate-2 from a driving clip Pose on Chroma only through img2img
Environments and sets The bases render sets well; Qwen-Image 2512 is strongest on signage and text Nothing keeps one set consistent across shots (new angles, next shot), nothing relights a subject into a plate
Costumes General instruction edits (klein, Qwen Edit 2511, HiDream edit) No try-on, outfit extraction or cosplay transfer
Movement and dance Wan 2.2 I2V/T2V with hip sway and hip bump pairs, Animate-2, SCAIL-2, VACE, LTX-2.5 (SFW) Nothing driven by music; no audio-driven singing or lip-sync on an adult_ok model; camera is two tilt LoRAs
Prompting Nothing. libs/isis/prompt-engineering scores LLM prompt templates and no apps/isis code imports it; HiDream's refine step was left out as fail-silent No prompt enhancer and no image or video captioner
Restoration and speed SeedVR2 3B/7B, FlashVSR, RealESRGAN, GIMM-VFI, lightx2v 4-step Wan LoRAs, LTX distilled LoRA Qwen edit-control runs 20 steps (no Lightning row); Animate-2 base schedule and SCAIL-2 turbo both wait on a lightx2v I2V distill LoRA row (A.01.06, A.02.06)

The worker image already has the nodes for most of the gaps. The object-info snapshot lists TextGenerate, TextGenerateLTX2Prompt, LTXVGemmaEnhancePrompt, WanDancerEncodeAudio, WanDancerVideo, WanSoundImageToVideo (+Extend), WanHuMoImageToVideo, WanUni3CControlnetApply, WanCameraImageToVideo, EmptyQwenImageLayeredLatentImage, ConditioningSetMask, ColorTransfer, ImageCropToMask, CropByBBoxes, ImageCompositeMasked and SAM3_Detect. Every recommendation below except the Qwen-Image-Layered row is weights plus a catalog entry, with no image rebuild.

Tier 1 — clear gaps, licence-clean, nodes already in the image#

  1. Prompt assist (no new weights). Two halves, both visible and editable before submit, both fail loud (a refusal, an empty reply or malformed output is an error the Studio shows, never a silently substituted prompt — the failure mode that kept HiDream's refine step out):
    • Text rewrite per model family in the generation API, using each family's official template: Z-Image's pe.py (Tongyi-MAI Z-Image-Turbo space), Wan's prompt-extend system prompts, Qwen-Image's polish template; LTX-2.5 keeps its own LTXVGemmaEnhancePrompt node. Chroma gets natural-prose photography language (its catalog already asks for that).
    • Caption from an image or clip in the graph through core TextGenerate on an encoder already in the library: qwen_2.5_vl_7b (Qwen-Image family) or gemma4_e4b_it_bf16 (library-only since 2026-09-14). This is what the MiniMax H3 character-replacement v3 workflow does with Qwen3-VL: one caption for the reference identity (face, hair, clothing, materials), one for the driving clip's action and scene, fed to the video model. For Isis it feeds wan22-i2v, wan22-animate-2 and scail2-animate.
    • The Easy Krea 2 "prompt enhancer" is the same trick: core TextGenerate running the model's own Qwen3-VL-4B text encoder with an "expert prompt engineer" system prompt (faithfulness first, do not invent clothing or props). Z-Image's and klein's qwen_3_4b is a candidate for the same use.
    • Risk to measure first: an instruction model may refuse or sanitise artistic-nude prompts. Fix a small prompt set (nude yoga, figure study, music-video wardrobe) and record refusal and drift rates per backend before picking one.
  2. Wan-Dancer-14B plus the lightx2v I2V 480p distill LoRA. One reference image and a music track → a dance clip whose motion follows the music (global keyframe expert plans the routine, local expert refines). Genres trained: Chinese classical, K-pop, street, tap, Latin. Apache-2.0, released 2026-07-13, nodes in the image. Weights: upstream Wan-AI/Wan-Dancer-14B @ 85ce88dd… global_model.safetensors 34.49 GB and local_model.safetensors 34.49 GB; umt5 and CLIP vision H shared. Comfy-Org/Wan-Dancer ships only fp8_scaled files, which D-V5 rejects, so the first proof is whether the core loader accepts the upstream files (their dtype is unread; 34.49 GB for a 14B model suggests part of it is stored above bf16). The build validator runs on zero-byte shadow files and cannot answer that, so it takes one real load on a worker (cents); if the loader refuses, that is a precision decision for the user. The template's LoRA, Kijai/WanVideo_comfy Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank64_bf16.safetensors (0.74 GB), is an extraction from the Apache-2.0 lightx2v model in a repo with no licence tag — record that provenance in the register. The same row also unblocks the Animate-2 base schedule and SCAIL-2's turbo path.
  3. Wan2.2-S2V-14B for singing and performance. Audio plus a still → full-body or half-body performance with lip-sync; Apache-2.0, native nodes. Comfy-Org/Wan_2.2_ComfyUI_Repackaged @ c4f60d30… wan2.2_s2v_14B_bf16.safetensors 32.59 GB plus wav2vec2_large_english_fp16 0.63 GB (already listed in V.12.04, never pulled). Run a two-clip bake-off against HuMo-17B (Comfy-Org/HuMo_ComfyUI @ 3a5e6947…, humo_17B_fp16 34.11 GB plus whisper_large_v3_fp16 3.09 GB, Apache-2.0, native): HuMo takes subject from the image, outfit and set from the text and timing from the audio, which suits music videos where wardrobe changes per shot. Keep the winner, list the other. LTX-2.5 also does audio and video but is SFW-only under Lightricks' acceptable-use policy. LongCat-Video-Avatar (MIT) needs a wrapper pack and ships as 64 GB of Diffusers shards — not first. Wav2Lip is non-commercial.
  4. Uni3C camera control. Camera path from a reference clip (orbit, dolly, crane) applied to Wan; native WanUni3CControlnetApply since ComfyUI 0.29. ewrfcas/Uni3C controlnet.pth 3.99 GB, Apache-2.0 (a pickle; Kijai's 2.0 GB safetensors conversion sits in an untagged repo). Trained on Wan 2.1 I2V, so one live clip on the 2.2 experts decides whether it stays. Wan 2.2 Fun Camera (59.16 GB) is the fallback, not the first pick.
  5. Qwen-Image-Edit 2511 wardrobe and set pack (SFW lane, all small). For clothed plates, costume design, set coverage and relighting — never on nude subjects (Qwen-Image is sfw_only; its outputs carry no acceptable-use clause, so a set plate made here can feed an adult_ok workflow).
    • Native 2511: Multiple Angles (fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA, Apache-2.0, 0.30 GB; 96 camera poses for the same set or costume), Lightning 4/8-step (lightx2v/Qwen-Image-Edit-2511-Lightning, Apache-2.0, 0.85 GB bf16; also gives qwen-edit-control its missing speed row), Anyone Cosplay 2511 (Civitai 2435164 / 2738032, all four flags, 0.44 GB).
    • Trained on Edit-2509, so each needs one render on 2511 before it is offered: Clothes Try On (1940532 / 2196278, all four flags), Outfit Extractor (1940557 / 2196307, all four), Pose Transfer v2 (1959609 / 2221229, all four), Next Scene v2 (lovis93/next-scene-qwen-image-lora-2509, MIT), Relight and Light Migration (dx8152/Qwen-Image-Edit-2509-Relight, dx8152/Qwen-Edit-2509-Light-Migration, Apache-2.0).
  6. Z-Image look pack (service tier, tiny). AfterDark v2.0-ZIT (2176378 / 2470291, Image+Rent, trigger 4ft3rd4rk; low-key, noir and fashion lighting), GrainScape UltraReal ZIT (1332651 / 2460872, all four; analog film), Movie Wide Angle (2186574 / 2462010, all four; ultra-wide cinema framing), [ZIT] Light Slider (2262501 / 2546755, all four, 0.8 MB). Cinematic Shot ZIT (432586 / 2454927) is Image+RentCivit, so internal_only.
  7. A detailer tail for stills (no new weights). SAM3_Detect on "face" or "hands" → ImageCropToMask → inpaint the crop at higher resolution on the same checkpoint → ColorTransfer → ImageCompositeMasked. This is the pattern in Legendaer's workflows (face, eyes and hand detailer passes, colour match after the inpaint) built from core nodes, and it targets the hands and faces that complex yoga poses break.

Tier 2 — useful, after tier 1 or on a specific need#

  • Z-Image composition → Chroma refine, two passes in one latent. Z-Image and Chroma share ae.safetensors (same LFS oid, proved 2026-09-13), so a Z-Image pass with pose or depth control can be refined by a Chroma checkpoint at 0.3–0.4 denoise without a decode. That gives Chroma the pose control C.12.03 found missing. Idea from the Easy Krea 2 workflow's dual-checkpoint sampling.
  • Regional prompting (ConditioningSetMask, core): separate prompts for subject, wardrobe and background regions, as Legendaer's colour-mask regions do.
  • Qwen-Image-Layered (Comfy-Org/Qwen-Image-Layered_ComfyUI, Apache-2.0, 40.86 GB bf16): splits a set into RGBA layers for prop and background edits. Heavy; only if set editing becomes routine.
  • The Walk Wan 2.2 I2V pairs (1363473, all four flags, "Normal", "Slow" and "Lite" high/low pairs, 0.29 GB each): natural runway and street walks.
  • Retro Sci-fi 90s anime, Chroma version (912942 / 2326231, Image+Rent, 13.5 MB): the one link from the user's list that fits a base Isis runs and its flags; only if a stylised anime lane is wanted.
  • A driving-clip library for Animate-2 and SCAIL-2 (yoga flows, choreography, runway walks; own footage or cleared stock). For movement quality this beats any LoRA; it is the same asset the T.12 motion library builds.
Link Verdict
MiniMax H3 character replacement (2855941), H3 Ultra Fastest (2835250), H3 Cinematic Look (2908686) The MiniMax H3 Community License (2026-08-02) excludes the EU, UK, Korea and the US from using, hosting or distributing the weights or their outputs; Oshun Synthetics LTD is a UK company. Ideas kept: SAM mask plus VLM captions (tier 1 item 1), 4-step distill plus a video upscaler (Isis already has both).
Nipple Play – MiniMax H3 (2888071) Sex-act LoRA, and on an excluded base.
10eros_Max_Int8_ref2va (2868392) H3 base (excluded); int8 only (fails the no-quantization rule D-V5); a merge that grafts Krea 2 turbo weights, so Krea 2's licence rides along; the author calls it an early experimental build.
Anima (2458426), Anima Delta Mix (2898169), One Obsession Anima (1318945), MiaoMiao Harem (934764), Vixon's Milk Factory (1882087), Galena Blend 1.5 (16300) Anima's weights are CircleStone Labs Non-Commercial (outputs may be used commercially, but running the weights in a monetised product needs a separate licence); its own card says it "will not work well at realism". Delta Mix has no commercial flags; MiaoMiao and Vixon's are RentCivit only (no right to sell images); Galena 1.3–1.5 are paid.
Easy Krea 2 workflow (2790822), Moody Krea 2 Mix (2731187), NSFW LoRA Krea2/Flux (655753) Not on its first ground (corrected below in section 6): the licence does not ban adult nudity and has no seat cap. The real limits are a revocable licence (30 days' notice) and commercial rights, outputs included, only below $1M company revenue — a benchmark candidate, not a Lilith backbone. Moody's Premium version is paid. The workflow's prompt enhancer is learned (tier 1 item 1).
SNOFS (1972981) Sex-act LoRA; RentCivit only; Krea 2, klein 9B (non-commercial) and Qwen bases.
NSFW MASTER (667086) The Krea 2 Turbo version is paid; the free Z-Image Turbo V2 (Image+Rent) is genitalia-focused and duplicates BetterNudes v3 and Amateur Studio.
Add Micro Details (1377820) Versions for Krea 2, Anima, Illustrious, Pony, NoobAI, Flux 1 — none for a base Isis runs. Detail on Z-Image is covered by the Detail Slider and DetailDeamonZ already in the library.
Zenith Z-Image Turbo (2237711) Every version is paid (Civitai paidAccess, permanent) and the flags are Image+RentCivit, so at best internal-only.
IntoRealism (1609320) Z-Image versions are fp8 only (D-V5) and the flags lack Image, so generated images could not be sold; the Krea 2 V4 version is sold on Ko-Fi.
One Obsession v24 (1318945), SteinIllustrious (1078146), Smooth Detailer Booster (1145743), Dynamic Poses slider (438059), WaduStyle (1019079), Ratatatat74 (107474), Galena Blend 1.4 / 1.3 (16300) Illustrious, NoobAI and Pony are a new SDXL-class base family (tag prompting, UNet, anime-first) whose photoreal output trails Chroma and Z-Image; adding one means a whole new pipeline line. NoobAI's licence forbids commercialising generated products; Illustrious is Fair AI Public License 1.0-SD. Per item: WaduStyle and Ratatatat74 lack Image (and Ratatatat74's newest trigger is "netorare"); Smooth Booster is Image+RentCivit and its Anima versions are paid; Galena 1.3/1.4 are paid; Dynamic Poses is a good idea with no version for our bases.
Demo AI Influencer 014 (1853522) Someone else's identity LoRA, a demo for a paid custom-LoRA service (24 downloads), Image+RentCivit. Isis characters should be trained on Isis's own consented references.
Retro Sci-fi 90s anime (912942) The exception: Chroma and Z-Image versions exist with Image+Rent. Tier 2, only for a stylised lane.
ComfyUI Image Workflows, Legendaer (1386234) SDXL-only graphs, but worth reading: detailer passes, colour match after inpaint, FBCNN JPEG clean-up of input photos, colour-mask regional prompting, wildcards, TIPO (a tag expander for SDXL, not useful for T5 or Qwen encoders). Carried into tier 1 item 7 and tier 2.

No link is in Civitai early access today (every earlyAccessConfig is null); the paid ones are Zenith, Galena Blend 1.3–1.5, Smooth Booster's Anima versions, Moody Krea 2 Premium and NSFW MASTER's Krea 2 Turbo.

4. Costs and space#

  • Library: 813.38 GB of the 900 GB alarm line after the arsenal pull. Tier 1 adds about 70 GB (Wan-Dancer with its LoRA), 33 GB (S2V) or 37 GB (HuMo), 4 GB (Uni3C) and about 3 GB of LoRAs. That fits only after A.05.03's retirements (LTX-2.3 80.38 GB, Animate-1 35.99 GB, Depth-Anything-V2 1.34 GB) bring the library to about 695 GB; the unused Animate-2 base (32.79 GB, library-only) and Gemma 4 E4B (if not used for captions) are the next space if needed.
  • Cache: each video family (Dancer ~70 GB, S2V ~33 GB) fits the 188 GB rotating space beside the pinned core.
  • Proof renders: one clip per new video model at the V.11 minimum shape is roughly $0.10–0.35 each on the video endpoint; stills and LoRAs $0.003–0.07 each. About $3–4 for all of tier 1, asked before spending as A.00.05 does.

5. A note on rows already in the library#

Given the boundary above, several rows approved on 2026-09-13 lean sexual rather than artistic: the internal-only Wan pairs Bouncing Boobs, Twerking, Slow Twerk and NSFW Posing Nude, the two breast sliders, and the explicit-rated UnGloryHail and NSFW Realism options. They are internal-only or policy-gated and need no action; they are listed in case the user wants to prune them.

6. Licence-limited models that may beat the stack (added the same day)#

Decision later on 2026-09-14: none of these is taken. The user dropped every model that would enter only for benchmarking under a licence not good for production, and any outreach emails; the approved scope is tracker sections E.00–E.08 in ISIS_CHROMA_RUNPOD_MVP_TODOS_2026-09-11.md.

The user asked which excluded models may be better than what Isis runs, to keep them for internal benchmarks. Licence texts were read verbatim, not from summaries.

Model Likely better at What the licence allows for an internal benchmark
MiniMax H3 (33.1B, video with stereo audio) Complex action, reference-driven camera and multi-reference control (independent comparisons, August 2026) Nothing without authorisation. The agreement's scope "is expressly limited to the Applicable Territory" and accepting it covers "using, reproducing, modifying, distributing, running, or displaying" the works; the UK, EU, US and Korea are Excluded Territories, so a UK company running it anywhere, including RunPod EU-RO-1, has no grant. The licence invites a request ("you are welcome to contact us about obtaining a license … based on robust controls and guardrails") to api@minimax.io. Its acceptable-use exhibit has no adult-nudity ban. MiniMax's hosted API is available in every region.
Krea 2 Raw / Turbo (12B) Identity preservation and style range; realism against Z-Image is contested (one 295-prompt test gives it to Z-Image) Allowed. Community Licence v1 (2026-06-22): commercial use of the model, derivatives and outputs only below USD 1M company-wide trailing revenue including affiliates; §4.2 requires filters against "prohibited, harmful, or unlawful content through your deployment", and the acceptable-use policy prohibits CSAM, non-consensual intimate imagery and likeness violations — not consensual adult nudity; no seat cap appears in the text; the licence is revocable and §9.2 lets Krea terminate "for any reason" on 30 days' notice. The 2026-09-13 exclusion over-read §4.2 and cited a seat cap the text does not contain.
FLUX.2 klein 9B (and the ponpoke uncensored encoder) Quality over the klein 4B Isis runs Allowed. The non-commercial licence counts "use by commercial or for-profit entities for testing, evaluation, or non-commercial research and development in a non-production environment" as non-commercial; revenue-generating use, anything that affects end users, and training or distilling models for commercial use are not.
Anima Anime and illustration only; not better for photoreal work Allowed, same wording as FLUX (CircleStone Labs Non-Commercial Licence). Low value unless an anime lane is planned.
Tencent Hunyuan (Video 1.5, Image 3.0, 3D) Unclear against Wan 2.2 and Z-Image; Hunyuan3D matters to the T track, not here Same territory exclusion as H3 (EU, UK, Korea): no grant.
Sulphur-2 — No licence at all, so no permission, benchmark included.
Closed APIs (Seedance 2.5, Wan 3.0, Qwen-Image 3.0 Pro, H3 Max) Probably all of the above Benchmarkable per call under each vendor's terms, SFW prompts only (the hosted services refuse nudity). @isis/seedance-provider already exists.

Three rules any benchmark-only tier needs, because the licences that allow evaluation also forbid what usually follows it: (1) a distribution value stricter than internal_only (which lets the operator render Oshun's own media), refused for every production job, operator included; (2) benchmark outputs kept in their own prefix and never used as LoRA or distillation data (FLUX and Anima count training for commercial use as commercial; H3's derivative definition covers training on synthetic outputs); (3) the licence register records the eval-only grant per row.

7. A stylised and anime lane (added the same day)#

The live RunPod catalog is realism-only. The older sacred-art, composition and flagship graphs are stylised but load SDXL / RealVisXL files that are not in the library and carry no content policy, so they are not a lane today.

Style LoRAs on bases Isis already runs (cheapest; no new base): Velvet's Mythic Fantasy Styles (599757; ZiT versions Detailed Anime 2504197, Anime Lines 2466747, Gothic Lines 2924569, Realistic Fantasy 2547883; all four flags; 238k downloads), Anime Flat Style (1952560; Qwen 2512 2559263 and ZiT 2455064; all four), Flat Color Style (1132089; Z-Image base 2637093, ZiT 2616765, Qwen 2143914, HiDream 1734377; Image+Rent), Character Design Sheet (100435 / 2500612, ZiT turnarounds with palette; all four), Dark Art Style (1380736; Chroma 2320313, ZiT 2467722; Image+Rent), Retro Sci-fi 90s (912942; Chroma 2326231, ZiT 2463751), Aesthetic Quality Modifiers (929497; ZiT 2622701, Qwen 2241189; Image+Rent), LTX-2.5 3D Animation (2895989 / 3273961; Image+Rent; SFW lane). Leave out LoRAs that imitate a named studio or living artist (Definitive Disney Studios, Afrobull, D-ART) — rights of publicity and trade-mark exposure, not licence flags.

Dedicated anime bases (Hugging Face licence tags read 2026-09-14):

Base Licence Verdict
Anima (4.18 GB) CircleStone Non-Commercial; outputs usable commercially; for-profit evaluation allowed; commercial licence by request Best quality in a June 2026 comparison (Anima > NewBie ≥ NetaYume > RouWei-Gemma >> SDXL). Eval-only now; ask CircleStone for a commercial licence.
NetaYume Lumina v4 (AIO 10.62 GB) Apache-2.0 (Lumina Image 2.0 base; lumina2 is a CLIPLoader type in the worker snapshot) Best licence-clean anime base; painterly look. First candidate for production.
Illustrious XL v2.0 CreativeML OpenRAIL-M (v1.1 SDXL licence; only v0.1 was Fair AI Public License) Commercially usable base with the largest LoRA ecosystem, but SDXL-era quality, and most Civitai Illustrious merges carry NoobAI lineage.
NoobAI-XL Fair AI Public License 1.0-SD plus a no-commercialisation term Out.
Pony V7 (AuraFlow, 7B) Pony licence: no inference service or application, no company above $1M revenue, no professional video production Out for Lilith.
NewBie-image Exp0.1 newbie-nc-1.0 Eval-only at most.

Video: Wan 2.2 I2V animates a stylised still natively; SCAIL-2 and Animate-2 can drive a stylised character from real footage (identity hold on a drawn reference is unmeasured).

One compliance gate before stylised nudity: anime styles skew youthful, and UK law (Coroners and Justice Act 2009, s.62) criminalises non-photographic sexual images of children, drawn ones included. An anime lane that allows nudity needs an adult-appearance check (prompt screening plus an output classifier) before it is enabled, not after.

Sources#