Anima 3.8B z3v0.1 is a style merge on LyloGummy's 52-block Anima 3.8B (the BASE / preview release, which pairs the DiT with a Qwen3.5 4B cross-attention adapter). Six Anima LoRAs — five style LoRAs plus an eye-detail LoRA — were remapped from 28 to 52 blocks and baked in at fixed weights, then the model was quantized to MXFP8. It is the 3.8B sibling of Anima 2B z3v0.1 and Anima 2.9B z3v0.1 and carries the same five styles, tuned for this architecture.

Compared with the 2B and 2.9B versions it follows longer natural-language prompts and multi-subject instructions better, and renders cloth and hair more realistically; the line art is softer.

Requirements (read first)

This is not a plain Anima checkpoint. You need all of:

• ComfyUI 0.33.1 or newer (52-block support and MXFP8 quantized layers; a GPU with MXFP8/FP8 support is recommended — tested on an RTX 5080)

• Custom node comfyui-anima-3-8B (the original v1 / preview node for Anima 3.8B BASE — not the v1.1 bundled-connector release)

• Anima-3.8B-expanded_adapter.safetensors (the Qwen3.5 adapter from the Anima 3.8B BASE release) in models/text_encoders

• qwen35_4b.safetensors (Qwen3.5 4B text encoder) in models/text_encoders

• qwen_3_06b_base.safetensors (native Anima text encoder) in models/text_encoders

• qwen_image_vae.safetensors in models/vae

• About 12–13 GB of VRAM at 896x1152 in our tests

Base model and adapter: https://civarchive.com/models/2880480?modelVersionId=3255578 and https://huggingface.co/lylogummy/Anima-3.8B . Qwen3 0.6B and the VAE: https://huggingface.co/circlestone-labs/Anima .

Recommended settings

• Steps: 30

• CFG: 4.5 (negative prompts work)

• Sampler / scheduler: res_multistep / beta

• Model sampling: none (no ModelSamplingAuraFlow node)

• Resolution: about 1 megapixel, e.g. 896x1152 (portrait), 1152x896 (landscape), 832x1216

• Do not add a turbo/distill LoRA and do not load the baked LoRAs again. Ordinary 28-block Anima LoRAs must be remapped to 52 blocks before they work.

Prompting

• No trigger words needed. Danbooru-style tags plus a short plain-English paragraph (scene, light source, mood) work best; the showcase prompts use exactly this pattern.

• Tip: start the prompt with enchanting eyes — it activates the baked eye-detail LoRA (richer irises and lashes).

• For several characters, give each one its own sentence and use names instead of pronouns.

• Close and half-body framing gives the sharpest faces; in wide shots faces get soft.

• Add shiny skin, oily skin to the negative if skin gets too glossy; adult woman, mature female plus child, loli, chibi in the negative for adult proportions.

• For full-body outfits, state the skirt length; hair tends to come out voluminous.

ComfyUI

UNETLoader (file in models/diffusion_models, weight dtype default) -> sampler. Text: CLIPLoader with qwen_3_06b_base.safetensors (type stable_diffusion) + Load Qwen3.5 4B (Anima) with qwen35_4b.safetensors + the Anima 3.8B adapter, all into the Anima 3.8B unified prompt node; use its expanded output as conditioning. VAELoader with qwen_image_vae.safetensors. Do not set the UNET loader to fp8 (the adapter path will crash); the MXFP8 layers load per-layer automatically.

Baked components (credits)

• Anima 3.8B, BASE (base) by LyloGummy, weight 1.0: https://civarchive.com/models/2880480?modelVersionId=3255578

• ushikani_kassen-styles, anima-base1.0 by zhihu, weight 0.6: https://civarchive.com/models/2760349?modelVersionId=3106457

• 448 - guanren79145223's Style, ANIMA by monkewt, weight 0.5: https://civarchive.com/models/2765265?modelVersionId=3112478

• CowCowCrab / Saiou-ga-ushi Style, v4.0 (Anima) by yeetyah145, weight 0.5: https://civarchive.com/models/1757934?modelVersionId=3165356

• @zxcvnmweruo Style Anima, v1.0 by Umbre99, weight 0.7: https://civarchive.com/models/2830588?modelVersionId=3193979

• [Anima]ppolar, v1.0 by ScarletDimension, weight 0.7: https://civarchive.com/models/2705088?modelVersionId=3037966

• Enchanting Eyes, Anima by EmilyMars, weight 0.8: https://civarchive.com/models/974076?modelVersionId=3027111

All LoRAs were remapped 28->52 blocks (existing blocks only; the 24 inserted blocks were left untouched). Thanks to the authors above, to LyloGummy for Anima 3.8B, and to CircleStone Labs for Anima.

License and use

Non-commercial. Derivative of Anima (CircleStone Labs Non-Commercial License), of Anima 2.9B / Anima 3.8B, and of the components above; their terms apply. Do not sell the model or offer it as a paid generation service.

中文简介

Anima 3.8B z3v0.1 是基于 LyloGummy 的 52 层 Anima 3.8B（BASE / 预览版，配 Qwen3.5 4B 适配器）的画风融合模型：把五个画风 LoRA 和 Enchanting Eyes 眼部 LoRA 从 28 层重映射到 52 层后按固定权重烤入，再量化为 MXFP8。与 Anima 2B / 2.9B z3v0.1 同一套画风，更听长句和多人描述，布料和头发更写实，线条偏软。需要 ComfyUI 0.33.1+、comfyui-anima-3-8B（v1 预览版节点）、Anima 3.8B 适配器、qwen35_4b、qwen_3_06b_base 和 qwen_image_vae。推荐 30 步、CFG 4.5、res_multistep + beta、约 100 万像素，不加 shift。不需要触发词；提示词开头加 enchanting eyes 可以让眼睛更细致。仅限非商用。