Langzaigg/moss_anime_turbo_merged

MOSS Anime Turbo Merged

中文说明

An unofficial anime-oriented diffusion model repository derived from OpenMOSS-Team/MOSS-Image-SFT. It contains the merged diffusion transformer (DiT) in BF16 and lower-memory INT8 ConvRot variants. It is not an OpenMOSS release and not an all-in-one checkpoint.

The fused-QKV backbone merges these community components into the official SFT weights:

• moss-cute-anime DoRA at 0.3 strength
• MOS-SFT 4-step turbo LoRA at 1.0 strength
• moss_pussy_lora_v2.0 at 0.7 strength

Repository samples are SFW. No additional LoRA or QKV-remapping patch is needed for either DiT variant.

Files and requirements

Two interchangeable DiT files are included. Use the BF16 file by default; use INT8 ConvRot when DiT memory or download size matters more than maximum precision.

Variant	File	Size	SHA-256
BF16	moss_anime_bf16_diffusion.safetensors	10,164,655,840 bytes (9.47 GiB)	0d6d92e580fa0611df08eca11e5250d5783b8fa6730a7e60175803ad79a42301
INT8 ConvRot	moss_anime_int8_convrot_diffusion.safetensors	5,413,692,384 bytes (5.04 GiB)	33b1daa4d25fb8f8e1ae38d43851162e78b8637f5735ae16f723626ab9e6648a

Release tested: 2026-10-26. The INT8 file stores 168 transformer-layer projection weights as row-wise INT8 ConvRot with FP32 scales; norms, embedders, modulation, and output heads remain BF16. It passed a ComfyUI 4-step smoke generation.

The text encoder and VAE are not included. Download or reuse the corresponding MOSS-Image components separately:

Component	Expected ComfyUI file	Directory	Loader
Diffusion model	moss_anime_bf16_diffusion.safetensors or moss_anime_int8_convrot_diffusion.safetensors	models/diffusion_models/	UNETLoader, default dtype
Qwen3.5-9B INT8 text encoder	mossImage_sft_txt_int8.safetensors	models/text_encoders/	CLIPLoader, type moss_image
FLUX.2 VAE	flux2-vae.safetensors	models/vae/	VAELoader

Use a current ComfyUI version with MOSS-Image support. This model is intended for the three-part UNETLoader + CLIPLoader + VAELoader workflow, not CheckpointLoaderSimple.

Recommended sampling

Parameter	Recommended default
Steps	4
Sampler	res_multistep
Scheduler	simple
CFG	1.0
Negative prompt	Empty
Denoise	1.0 for text-to-image
Resolution	Around 1 MP; width and height divisible by 16

euler + simple at 4 steps is a useful alternative when a seed has a local anomaly. The upstream turbo recommendation is 8 steps with euler + simple; the 4-step recommendation here comes from controlled tests of this derivative and should not be applied indiscriminately to the official SFT model.

MOSS-Image accepts natural-language captions and Layout JSON for positioned people, objects, panels, and literal text. Reference images can be supplied through the multimodal text encoder for semantic guidance; this does not guarantee identity preservation.

ComfyUI-AnimaFlow provides ready-to-use ComfyUI workflow templates and HTTP/MCP integration.

Evaluation

Nineteen text-to-image scenes were run with identical prompts, seeds, and resolutions on an Arc A770M 16 GB system. Both versions used res_multistep + simple, 4 steps, CFG 1.0, and an empty negative prompt.

Version	Median generation time	Mean generation time	Trimmed mean*
Previous AIO INT8 derivative	28.2 s	31.0 s	29.2 s
This BF16 model / three-part workflow	27.4 s	29.7 s	28.7 s

* Mean after excluding the two fastest and two slowest results. Timings cover successful generations only and are specific to this machine and mixed-resolution workload.

All 19/19 scenes eventually produced successful results, across 20 attempts. The autumn-scene attempt hit UR_RESULT_ERROR_DEVICE_LOST once and succeeded after ComfyUI restarted; this was not an uninterrupted crash-free run.

Manual side-by-side inspection rated this version better in 14 scenes, the previous version better in 1, and 4 broadly tied. Improvements were mostly in clothing, backgrounds, materials, and some object interactions. The speed difference is small and should be considered broadly comparable rather than a guaranteed speedup.

This comparison changed both the model weights and the loading workflow, so it does not isolate the effect of BF16 precision by itself. One no-artist scene regressed to a duplicate tail. Exact text, negative-concept suppression, complex anatomy, and reliable character reproduction remain unresolved. Image-editing and reference-image paths were outside this evaluation.

The INT8 ConvRot variant was verified with a successful 4-step ComfyUI generation but was not included in the 19-scene quality or timing comparison; evaluate it separately before using it for final production work.

Samples

All samples are SFW and use this BF16 model, res_multistep + simple, 4 steps, CFG 1.0, and no negative prompt.

Picnic	Kitchen

Autumn	Sketchbook

Limitations

• Complex anatomy, object interactions, exact layout adherence, and negative-concept suppression can fail.
• Small or multi-paragraph text may be malformed; short literal strings work best.
• Reference images do not guarantee exact identity, clothing, or logo reproduction.
• On 16 GB GPUs, the BF16 DiT and large text encoder can create significant memory and paging pressure.

License and credits

Distributed under Apache-2.0; see LICENSE. This is a community merge, not an official OpenMOSS release. Review the licenses and terms of the base model and every merged community component before redistribution or production use.

• OpenMOSS-Team/MOSS-Image-SFT
• Community moss-cute-anime DoRA, merged at 0.3 strength
• Community MOS-SFT 4-step turbo LoRA, merged at 1.0 strength
• Community moss_pussy_lora_v2.0, merged at 0.7 strength

---

中文说明

这是基于 MOSS-Image-SFT 的非官方二次元合并 DiT 仓库，不是 OpenMOSS 官方发布。fused-QKV 主权重融合了 moss-cute-anime DoRA（0.3）、MOS-SFT 4 步 turbo LoRA（1.0）和 moss_pussy_lora_v2.0（0.7），无需再挂这些 LoRA 或 QKV 转换补丁。

文件与使用方法

本仓库只有扩散模型，不是 AIO 单文件，请另外准备：

• moss_anime_bf16_diffusion.safetensors → models/diffusion_models/（默认，完整 BF16）
• moss_anime_int8_convrot_diffusion.safetensors → models/diffusion_models/（低显存 / 小体积备选）
• mossImage_sft_txt_int8.safetensors → models/text_encoders/
• flux2-vae.safetensors → models/vae/

ComfyUI 使用 UNETLoader + CLIPLoader + VAELoader 三段式加载，CLIP 类型选 moss_image，不要使用 CheckpointLoaderSimple。可以直接使用 ComfyUI-AnimaFlow 的工作流、HTTP API 或 MCP 工具。

推荐默认：4 步、res_multistep + simple、CFG 1.0、空负面、文生图 denoise 1.0。分辨率约 1MP，宽高为 16 的倍数。某个种子出现局部异常时，可改用 euler + simple、4 步重试。

实测与限制

在 Arc A770M 16 GB 上，19 个同提示词、同种子、同分辨率场景中，最终 19/19 成功，共尝试 20 次；其中一次遇到 XPU DEVICE_LOST，重启后续跑成功。相对旧 AIO INT8 版本，人工目检为新版 14 项更好、旧版 1 项更好、4 项持平。中位耗时 27.4 秒对 28.2 秒，整体属于同档速度。INT8 ConvRot 变体已通过一次 ComfyUI 4 步生成实测，但未纳入这轮 19 场景质量横评。

这次横评同时改变了权重与工作流，不能单独归因于 BF16。复杂肢体、文字、负面概念抑制与角色精确还原仍可能失败，也出现了重复尾巴的局部回归。16GB 显存环境仍需注意文本编码器与 BF16 DiT 的换页压力。

仓库展示样例均为 SFW。