MOSS 4 ULTRA-REAL fêtesnap Snapshot style

紧急加更！MOSS 4 节日快照 超拟真 快照风格 内置加速 12 steps, CFG=1 Built-in DMD booster

FineView | 欢视 MOSS IMAGE DEEPSICK /Editor

MOSS-Image Merged 6vidit9`s DEEPSICK

基于 OpenMOSS MOSS-Image-SFT / Base 的融合衍生模型。两个官方变体按层带比例融合，并混入美学与写实权重；编辑能力与 DMD 加速均未注入底模，改由配套工作流旁路启用与平台侧挂载。

A merged derivative of OpenMOSS MOSS-Image-SFT / Base. The two official variants are merged band-by-band with aesthetic and realism weights mixed in. Editing and DMD acceleration are not baked into the base — they are enabled as a bypass in the companion workflow and mounted on the platform side.

一、模型定位与核心能力 / Core capabilities

底座 — MOSS-Image，5B 单流 DiT，42 层，128 通道 FLUX.2 latent。

Base model: MOSS-Image, a 5B single-stream DiT with 42 blocks and a 128-channel FLUX.2 latent.

本质 — 两个官方变体按层带比例融合，并混入NSFW、美学、写实三类权重，产出单一 checkpoint。
编辑权重与加速权重留在模型之外。

Essence: the two official variants merged band-by-band with high-resolution NSFW, aesthetic and realism weights mixed in, producing one checkpoint. The edit weights and the speed weights stay outside the model.

能力 — 文生图 · 原生布局控制 · 中英长文本渲染。指令式图像编辑由配套工作流以旁路方式提供，可自行开关。

Capabilities: text-to-image · native layout control · bilingual long-text rendering. Instruction editing is provided by the companion workflow as a bypass, switchable at will.

美学基调 — 本版的主要视觉语言由 Cinematic photography（6vidit9） 奠定：电影级光比、克制的色彩与胶片质感。再经 Mystic XXX（Alcaitiff）的 MOSS 版本完成艺术与性爱的交融——在保持画面质感的前提下放开成人表达。

The visual language of this edition comes from Cinematic photography by 6vidit9: cinematic contrast, restrained colour and a filmic finish. Mystic XXX by Alcaitiff, MOSS edition, then fuses artistry with eroticism — adult expression opened up without losing the surface quality.

推理 — 平台侧挂载 DMD 加速后 8-25 步、CFG 1，1.5 MP 单张约 12–18 秒（RTX PRO 6000）。未挂加速时 25 步、CFG 2.5。

Inference: 8-25 steps at CFG 1 with the DMD speed LoRA mounted on the platform, about 12–18 s per 1.5 MP image (RTX PRO 6000). Without acceleration, the official 25 steps at CFG 2.5.

许可 — Apache-2.0，派生自 Apache-2.0 上游权重，致谢名单见下节。

License: Apache-2.0, derived from Apache-2.0 upstream weights; full credits in the next section.

二、许可与致谢 / License and credits

License: 本模型为 Apache-2.0 上游的衍生作品，再分发时请保留上游声明并说明修改。

License: This is a derivative of Apache-2.0 upstream weights. Keep the upstream notices and state your changes when redistributing.

上游权重

• MOSS-Image Base / SFT — OpenMOSS-Team — Apache-2.0

• ComfyUI repack — Comfy-Org — Apache-2.0

• Text encoder Qwen3.5-9B — Qwen — Apache-2.0

• VAE FLUX.2-klein-base-4B — Black Forest Labs — Apache-2.0

美学基调（特别致谢）

• Cinematic photography — 6vidit9 — https://civarchive.com/models/2682289 — 本版美学的主要奠基 SFW

• [Moss] Mystic XXX — Alcaitiff — https://civarchive.com/models/3015391 — NSFW

本版的气质几乎全部来自 Cinematic photography。它的电影级布光与色彩逻辑被适配到 MOSS 后，成为底模的默认视觉语言；Mystic XXX 的 MOSS 版本在此基础上加入成人向的表现力，二者共同构成这一版的辨识度。

The character of this edition is almost entirely Cinematic photography: its cinematic lighting and colour logic, adapted to MOSS, became the model's default visual language. Mystic XXX's MOSS edition adds adult expressiveness on top, and the two together define what this version looks like.

重要权重

• Hi-Res LoHa v0.5 — comfyanonymous — Apache-2.0

• [Moss] Realism Enhancer — Alcaitiff — https://civarchive.com/models/3023438 — SFW

• Amateur nudes（MOSS version） — konflikti — https://civarchive.com/models/764517 — NSFW

• INT8 ConvRot quantization — toxicdog — Apache-2.0

保留致谢

• 8-step DMD distillation — lodestones — Apache-2.0 — 平台生成器已支持加速 LoRA 挂载

• Edit LoRA armV5 — akhaliq / MOSS-Image-Edit-HQ — https://huggingface.co/akhaliq/MOSS-Image-Edit-HQ — Apache-2.0 — 由配套工作流旁路启用

• LCS colour / sharpness — facok / ComfyUI-LCS — https://github.com/facok/ComfyUI-LCS — MIT

• LCS 官方代码 — https://github.com/ExplainableML/LCS

• The Latent Color Subspace — Pach, Bader, Bouniot, Belongie, Akata — https://arxiv.org/abs/2603.12261 — ICML 2026

训练器及技术 Trainer & Technology

• ai-toolkit — ostris — https://github.com/ostris/ai-toolkit — Apache-2.0

The editing capability here comes directly from ai-toolkit. That project implements the moss_image architecture's edit mode; the armV5 weights and all of their ablation runs were trained on it, and its Edit Training layout (including the kv_cache option) is exactly what this model's edit_format refers to. Special thanks to ostris and the project.

本模型与上述任何一方无隶属、赞助或背书关系。

No affiliation, sponsorship, or endorsement by any of the above is claimed.

三、融合方案 / Merge scheme

3.1 组成 / Composition

代码块

diffusion    MOSS-Image-Base / SFT              band-wise merged
 resolution   Hi-Res LoHa v0.5                   merged
 aesthetic    Cinematic photography (MOSS)       merged   ← 美学主基调
 aesthetic    [Moss] Mystic XXX                  merged
 realism      [Moss] Realism Enhancer            merged
 snapshot     Amateur nudes (MOSS)               merged

 editing      Edit LoRA armV5                    bypass, in the workflow
 speed        8-step DMD distillation            mounted by the platform

3.2 分块比例与理由 / Band ratios and why

42 个 DiT block 按深度切成三段，每段独立比例。

The 42 DiT blocks are split into three depth bands, each with its own ratio.

代码块

Low     layers 0–20      ratio 0.5    composition, pose, layout
 Mid     layers 21–34     ratio 0.4    facial proportions, placement
 High    layers 35–41     ratio 0.7    texture, skin, hair, micro-detail

低中层的取值为 0.5 / 0.4，中层更偏 SFT 侧，承接蒸馏后的加速轨迹与五官比例；高层取 0.7 偏 Base 侧，因为高分辨率细节主要由深层 block 决定。分块同时避免了一个通病：单一比例对 42 层做同一个平均，而各来源在高频细节上最不一致，纹理总是最先被抹掉的——纹理恰恰是人像与写实的来源。

Low and mid sit at 0.5 / 0.4, the mid band leaning further toward SFT to carry the distilled trajectory and facial proportions; the high band leans to 0.7 toward Base, because fine detail lives in the deep blocks. It also avoids a common failure: a single global ratio treats all 42 blocks alike, and since the sources disagree most on high-frequency detail, texture is always the first thing averaged away — exactly what photorealism depends on.

3.3 同时具备两个官方变体的长处 / Strengths of both

Base 与 SFT 不是同一取向的两份权重：Base 的结构与细节保真更强，SFT 的成图完成度与指令响应更好。分块融合让二者共存于同一权重，而不是二选一。

Base and SFT are not two copies of the same thing: Base holds structure and detail better, while SFT finishes images and follows instructions better. Band-wise merging lets both live in one weight instead of forcing a choice.

3.4 融合精度 / Merge precision

所有权重的合并都在 fp32 下累加，最后一次性地降到 bf16 输出。不在 bf16 权重上逐层累加——社区 LoRA 的低秩增量通常只占权重 RMS 的 1–8%，而 bf16 在 |w|≈0.02 时的最小间隔是 7.8e-5，逐层在 bf16 上做加法会把大部分增量舍入掉。

All merges accumulate in fp32 and round to bf16 once at the end. Nothing is accumulated onto bf16 weights — a community LoRA's low-rank delta is typically 1–8% of the weight RMS, while bf16's spacing at |w|≈0.02 is 7.8e-5, so layer-by-layer addition in bf16 rounds most of the delta away.

融合顺序与精度必须在导出前确定，而不是靠推理时补。这是本模型与「在 bf16 权重上直接叠 LoRA」在效果上的主要差别。

Merge order and precision must be settled before export, not patched at inference. This is the main behavioural difference from simply stacking LoRAs onto bf16 weights.

3.5 为什么编辑与加速不进底模 / External Editor and DMD

编辑权重会干扰 SFT 的美学。 编辑 LoRA 训练于「指令 + 参考图」的条件分布，它重塑的是模型的响应方式；把这套权重平均进底模，会连带改变画面的色彩倾向与质感。因此编辑只作为旁路工具存在：需要时打开，不需要时关掉，底模本身不受影响。

Edit weights badly disturb the SFT aesthetic. An edit LoRA is trained on an instruction + reference distribution and reshapes how the model responds; averaging those weights into the base also shifts colour bias and surface quality.

加速同理。 平台生成器已经支持自定义挂载 DMD 蒸馏 LoRA，加速因此变成用户侧的运行时选择，而不是写死在权重里。

The same logic. The platform generator already supports mounting a DMD distillation LoRA, so speed becomes a runtime choice instead of a baked-in property.

四、推理链路 / Inference pipeline

代码块

UNETLoader (merged bf16)
    ↓
 LCS Sharpness Intervene          ← 采样期锐度子空间干预
    ↓
 SkipLayerGuidanceDiTSimple       ← single_layers = 22
    ↓
 BasicGuider ─┐
 BasicScheduler (simple, 25) ├→ SamplerCustomAdvanced → VAEDecode → Save
 KSamplerSelect (euler)      │
 RandomNoise                 ┘

本版目标即 25 步 / CFG 2.5。若确实需要 CFG（3.0–4.0），必须旁路加速权重并把步数提到 30 以上，否则颜色会漂移。

This edition targets 25 steps at CFG 1. If CFG (3.0–4.0) is genuinely needed, bypass the accelerated weights and raise steps past 30, otherwise colour drifts.

4.2 编辑分支 / ai-toolkit Editor

Text Encode MOSS-Image Edit（ComfyUI 0.43.0 原生节点）把指令与参考图编码为 conditioning。

The native Text Encode MOSS-Image Edit encodes instruction and reference into conditioning.

编辑能力通过配套工作流提供，默认关闭。工作流里是一个 LoraLoaderModelOnly + Text Encode MOSS-Image Edit 的分支，开关一个节点即可切换。

Editing ships with the companion workflow and is off by default: a branch with one LoraLoaderModelOnly plus Text Encode MOSS-Image Edit. Toggle a node to switch it on.

edit_format — 必须设为 ai-toolkit，与 armV5 的训练布局一致。选错不报错，但身份、构图、细节都会漂。

Must be ai-toolkit to match armV5's training layout. A wrong value does not error, but identity, framing and detail drift.

参考图尺寸 — 0.25 MP，即 512×512。VAE 为 16× 下采样且 DiT 不做 patchify（patch = 1），因此 512/16 = 32，参考序列共 1024 个 token，KV 约 0.5 GB。

Reference size: 0.25 MP, i.e. 512×512. The VAE downsamples by 16 and the DiT does not patchify (patch = 1), so 512/16 = 32 and the reference sequence is 1024 tokens, about 0.5 GB of KV.

参考图必须单独缩放，不要与输出尺寸共用缩放节点。参考 latent 会参与每一步注意力，用 2 MP 参考会直接打满显存。

Scale the reference separately; never share the output scale node with it. Reference latents attend at every step — a 2 MP reference will exhaust VRAM.

4.3 美学与写实的来源 / Where the look comes from

Cinematic photography（6vidit9）— 美学主基调。 本版的光比、色彩克制与胶片质感都由它奠定。它把「电影摄影」的布光逻辑带进 MOSS：主光方向明确、暗部保留层次、高光不过曝，画面天然带叙事感。适配为 MOSS 权重后，它成为底模的默认视觉语言。

The primary aesthetic. Contrast ratios, restrained colour and the filmic finish all come from here. It brings a cinematographer's lighting logic into MOSS — a clear key light, shadow detail preserved, highlights held — so images read as narrative frames. Adapted to MOSS weights, it became the model's default visual language.

[Moss] Mystic XXX（Alcaitiff）— 艺术与性爱的交融。 在电影感基调之上加入成人向的表现力：肌肤质感、身体比例与情欲张力。它没有覆盖掉前者的光色逻辑，而是让同一种质地同时承载叙事与欲望。

Adds adult expressiveness on top of that cinematic base: skin quality, body proportion and erotic tension. It does not override the lighting logic above; it lets the same surface carry both narrative and desire.

[Moss] Realism Enhancer（Alcaitiff）— 写实还原。 肤质、光线衰减、镜头质感，把画面从渲染感拉回照片感。

Realism restoration: skin texture, light falloff and lens character.

4.4 显存配置 / VRAM

代码块

merged model (bf16)      GPU     ~9.5 GB    ComfyUI auto streamed by layer
 text encoder (9B int8)   GPU     ~9 GB      ComfyUI auto offloaded between stages
 VAE                      GPU     ~0.3 GB
 reference KV (512²)      ~0.5 GB    cached once, held through sampling
 1.5 MP activations       GPU     ~2 GB

文本编码器与扩散模型由 ComfyUI 自动装卸，二者峰值不叠加。

The text encoder and the diffusion model are loaded and unloaded automatically by ComfyUI; their peaks do not stack.

五、推荐参数 / Settings

挂载 DMD 时（推荐）

代码块

steps             20
 cfg               1.0            (BasicGuider, no negative prompt)
 sampler           euler
 scheduler         simple
 shift             5
 skip layer guide  single_layers 22
 resolution        1.0 – 1.5 MP
 reference (edit)  512 × 512      (1024 tokens)

未挂 DMD 时（官方配方）

代码块

steps             25
 cfg               2.5            (CFG with negative)
 sampler           euler
 scheduler         sgm_uniform

文生图与编辑使用同一套参数；编辑时输出分辨率跟随参考图。

The same settings cover both text-to-image and editing. When editing, output resolution follows the reference.

25 步下编辑强度主要由步数本身提供，image_strength 用于微调贴合度。

At 25 steps most edit strength comes from the step count itself; use image_strength for fine adherence.

六、安装与运行环境 / Runtime

6.1 ComfyUI

需要 ComfyUI 0.43.0 或更高。Text Encode MOSS-Image Edit 是 0.43.0 引入的原生节点，低版本不会出现该节点，也无法按编辑布局构造 conditioning。

Requires ComfyUI 0.43.0 or newer. Text Encode MOSS-Image Edit is a native node introduced in 0.43.0; older builds neither show it nor build the edit conditioning correctly.

暂时无法升级时，可改用第三方插件 Chenkin-x/ComfyUI-Moss-Image-Edit：克隆到 custom_nodes/，权重放 models/model_patches/。该插件对版本敏感，与核心版本不匹配时会报 pe_embedder 错误，官方推荐仍是原生路径。

If upgrading is not possible, the third-party plugin Chenkin-x/ComfyUI-Moss-Image-Edit is an alternative: clone into custom_nodes/, put weights in models/model_patches/. It is version-sensitive and raises a pe_embedder error when mismatched, so the native path remains the recommended one.

6.2 必需文件 / Required

代码块

models/diffusion_models/   fineview_moss_v3.safetensors
 models/text_encoders/      qwen3.5_9b_int8_convrot.safetensors
 models/vae/                flux2-vae.safetensors

模型本体已包含编辑权重，无需额外加载任何 LoRA。文本编码器与 VAE 使用 Comfy-Org 的官方重打包版本，Quantization 格式须与 ComfyUI 原生支持的一致（int8 ConvRot 或 bf16）。

The checkpoint already contains the edit weights — no extra LoRA is needed. The text encoder and VAE are Comfy-Org's official repacks; their format must match what ComfyUI loads natively (int8 ConvRot or bf16).

6.3 工作流接线 / Workflow

代码块

Load Diffusion Model
    ↓
 LCS Sharpness Intervene (可选)
    ↓
 Skip Layer Guidance DiT (single_layers = 22)
    ↓
 BasicGuider ─┐
 BasicScheduler (simple, 25) ├→ Sampler Custom Advanced → VAE Decode → Save
 KSamplerSelect (euler)      │
 RandomNoise                 ┘

编辑分支另接：

代码块

Load Image
    ↓
 Image Scale To Total Pixels (0.25 MP / 512²)
    ↓
 Text Encode MOSS-Image Edit   ← 同时接 VAE
    ↓
 BasicGuider.conditioning

工作流内已接好两条可开关的分支：

• 编辑分支（默认旁路）：LoraLoaderModelOnly（armV5）+ Text Encode MOSS-Image Edit

• 加速分支（按平台能力挂载）：8-step DMD 蒸馏 LoRA

The workflow ships two switchable branches: the edit branch (bypassed by default) and the speed branch (mounted per platform support).

七、已知限制 / Known Limit

NSFW 内容 — 模型具备成人内容生成能力，使用范围受 Civitai 条款与各地法律约束，责任在使用者。25 步下细小手部与解剖结构仍可能出错。

Adult content: the model can generate NSFW material, bounded by Civitai's terms and local law, with responsibility on the user. Small hands and anatomy can still fail at 25 steps.

美学倾向强烈 — 本版以电影感与成人向表达为核心取向，画面天然偏暗部、偏克制。要求明亮通透或卡通风格的场景，建议换用 MOSS 官方权重。

A strong aesthetic bias: this edition is built around cinematic contrast and adult expression, so images run darker and more restrained. For bright, airy or cartoon looks, use the official MOSS weights instead.

编辑不在底模内 — 需要编辑时必须启用配套工作流的旁路分支；直接加载底模做编辑不会生效。

Editing is not in the base: it requires enabling the bypass branch in the companion workflow. Loading the checkpoint alone will not edit.

加速不在底模内 — 8-25 步的配置依赖平台侧挂载 DMD；未挂载时请按 25 步 CFG 2.5-3 使用，否则出图会偏软。

Speed is not in the base: the 8-25-step setting depends on the platform mounting DMD. Without it, use 25 steps at CFG 2.5-3 or images come out soft.

文本编码器 — 与扩散模型交替装卸，首次编码存在加载等待。

Text encoder: swapped in and out against the diffusion model, so the first encode waits on a load.

LCS 首次校准 — LCS 组件第一次运行时会分析 VAE 权重，建立颜色子空间与锐度子空间的基与锚点，产生一次性的校准耗时。结果按 VAE 指纹缓存，其后直接复用；更换 VAE 后需重新校准。

LCS calibration: on its first run the LCS nodes analyze the VAE weights to build the basis and anchors of the colour and sharpness subspaces, costing a one-time calibration pass. The result is cached against the VAE fingerprint and reused afterwards; a different VAE needs recalibration.

参考图预算 — 编辑权重训练时的参考预算约为 256 token（约 256×256）。512×512 参考是 1024 token，为训练规模的 4 倍，已足够覆盖构图与身份；继续提高到 1024×1024 只多换细节，代价是速度与显存，且更偏离训练分布。

Reference budget: the edit weights were trained around a 256-token reference (~256×256). A 512×512 reference is 1024 tokens, 4× the training size — enough to carry composition and identity. Going to 1024×1024 buys detail only, paying in speed and VRAM, and drifts further from the training distribution.

多参考图 — 架构支持至多 4 张，但 Editor 模型以单参考训练。

Multi-reference: accepts up to 4 references, but was trained on single-reference pairs.

八、速查 / Quick check

代码块

merge        INT8 ConvRot → BF16 dequant → band-wise lerp → BF16 output
 bands        band1_end = 21 , band2_end = 35
 high ratio   0.7  (leans to the Base / detail side)
 baked        Hi-Res LoHa, Mystic XXX, Realism Enhancer, Amateur nudes, armV5 edit LoRA
 precision    fp32 accumulation, single bf16 rounding at export
 rope.bands   kept FP32 throughout (positional precision)
 sampling     CFG 2.5, 25 steps, euler + simple
 reference    512×512 = 1024 latent tokens (KV ≈ 0.5 GB)
 sharpness    LCS sharpness subspace, acts during sampling
 content      photographic realism, NSFW-capable 

文档版本 / Revision: v1.3 · 2026-10-29

FineView | 欢视 MOSS 2 ASIAN 东亚

with RED-MOSS Merged acceleration v2.1

EULER / 6-8 Steps Simple / CFG=1 / LCS (Latent Color Subspace)

版本说明（Notes）：
该版本修复了 “RED-MOSS 1” 由于加速蒸馏痕迹过重，人物皮肤不好的问题。
Fixes the issue with "RED-MOSS 1" where the skin looked bad due to excessive acceleration distillation effects.

商用声明（Commercial Statement）：
MOSS 2 ASIAN 东亚 版本在 RED-MOSS 1 Merged acceleration 基础上进行了新的分层方案
并且通过引用 Moss Image - Better Nude 1.0 by @salttaro 在6步基础上获得了更好的东亚特征与Nude表现。
所以，该版本将会把 50% BUZZ 收入通过站内TIPS方式，回馈给贡献者。

为了便于监督，该版本将开放下载量展现与BUZZ⚡️收入展现。

为什么会有皮肤问题 Why does skin issues happen?

1. 频率分布不同。(Different frequency distributions. Images and text have high contrast, low to mid-frequency content: sharp edges, flat areas, so the error from 6 steps hardly shows up in these areas. A human face is dominated by mid-frequency features (skin tone transitions, the gradient from cheekbones to jawline, eye socket shadows), which are precisely the parts of the spectrum that are lost first in a few-step distillation. Text being "perfect" doesn't mean the model is good—it just means text isn't sensitive to this kind of error. ) 图表和文字是高对比、低中频的内容：边缘锐利、区域平坦，6 步的轨迹误差落在这些地方几乎不显现。人像脸是中频主导（皮肤明暗过渡、颧骨到下颌的渐变、眼窝阴影），恰恰是少数步蒸馏最先丢掉的那一段频谱。文字"完美"不代表模型好，只代表文字对这种误差不敏感。

2. 潜在分辨率摊薄。( Potential resolution dilution. FLUX.2 VAE uses 16× downsampling. In an 1MP image, a face might only occupy 8×10 latent patches—so the error from each of the 6 steps is spread across so few patches that the face ends up looking "plastic." 2D faces are already represented by lines and color blocks, so they don't need these gradients and are therefore unaffected. ) FLUX.2 VAE 是 16× 下采样。1MP 分辨率的图里，一张脸可能只占 8×10 个 latent patch——6 步里每一步的误差平均到这么少的 patch 上，脸就糊成"塑料感"。二次元脸本来靠线条和色块表达，不需要这些渐变，所以不受影响。

3. 三路合一时的平均效应。 (The averaging effect when combining the three paths. The objective functions of the three paths you're merging are different (direct regression / DMD / average flow). Any weighted average will prioritize erasing high-frequency details because high frequencies are the most inconsistent among the three. And the face is the area with the densest high-frequency changes in the image.) 合并的三条路径目标函数不同（直接回归 / DMD / 平均流）。任何加权平均都会优先抹掉高频细节，因为高频在三者间最不一致。而人脸是画面里高频变化最密集的区域。

如何修复 How do I fix it ？

经过高效的 Agents+VLM 分层测试：

• 低层 ratio_low（0–20）：构图、姿态、大结构 Composition, pose, overall structure

• 中层 ratio_mid（21–34）：五官位置与比例 Facial features placement and proportion

• 高层 ratio_high（35–41）：纹理、皮肤、发丝细节 Texture, skin, and hair details

高层携带的正是皮肤渐变、发丝边缘这些高频信息，而三条路径在高层的一致性最差——

一平均就抹平，人脸最先崩，而文字和图表因为依赖边缘对比、不依赖中频过渡，所以看起来没事。分块之后，你可以让高层偏向细节保留最好的那条路径，低层仍保持构图稳定。The high-level features carry high-frequency information like skin gradients and hair edges, but the consistency among these three paths is the worst at high levels—

When averaged out, everything gets smoothed out, and faces are the first to suffer. Text and charts, on the other hand, rely on edge contrast rather than mid-frequency transitions, so they look fine. After segmentation, you can let the high-level paths retain the best details while keeping the low-level structure stable.

以上信息仅代表 RED-MOSS 1 Merged acceleration v2 方案进行推导，不代表所有MOSS版本。

RED-MOSS 1 Merged acceleration

》REDMOSS-图片-T2I-LCS-6STEPS_v1 10/19 付费工作流 ⚡ Pay/Access WF

LCS (Latent Color Subspace) nodes for in-sampling color and sharpness control.

• A 自训练蒸馏 LoRA（rank 128，覆盖 42 层）固定 4 步直接回归真实 latent
  Euler + simple，CFG 1，4 步 （最终版本为6-8步）

• B lodestones 8-step DMD整模型级权重data-space DMD 蒸馏Euler + simple，CFG 1，8 步

• C congruency MFD（Base）LoRA（约 72 MB）Mean Flow Distillation 4–6 步

• 未加速的官方权重仍适用其原生参数：50 步、CFG 3、sgm_uniform、shift 5 （Unaccelerated official weights keep their native recipe: 50 steps, CFG 3, sgm_uniform, shift 5.）

技术说明 / Technical note: 三条路线的训练目标不同（A 为无教师直接回归，B 为 DMD 分布匹配， C 为平均流蒸馏），因此不保证可无损叠加；任何叠加结果均属新的独立派生作品，不在原作者质量担保范围内。

The three routes use different training objectives (teacher-free regression, DMD distribution matching, mean-flow distillation) and are not guaranteed composable without loss. Any combination is a new, independent derivative outside the original authors' quality guarantees.

Commercial Use & Modification Notice v2.0 — MOSS-Image Derivative

本文件依据 Apache License 2.0 §4（保留声明、标注修改）与 §6（商标）编写， 适用于本项目产出的全部 MOSS-Image 衍生权重、LoRA、量化模型、融合模型与打包 checkpoint。

Issued under Apache License 2.0 §4 (retain notices, state changes) and §6 (trademarks). It covers every MOSS-Image derivative produced by this project.

0. 版本记录 / Version history

版本日期变更v1.02026-10-18初版：单一加速路径（自训练蒸馏 LoRA）+ Base/SFT 混用v2.02026-10-18合并三条加速路线；新增融合权重、量化重打包、AIO 打包说明；v1.0 全文保留于附录 A

v1.0 remains in force for artifacts published before this revision and is reproduced verbatim in Appendix A. v2.0 supersedes it for all artifacts published after.

1. 基础组件与许可 / Base components and licenses

全部衍生作品建立在下列 Apache-2.0 组件之上。原始版权归各自作者所有，本项目不主张任何权利。All derivative works build on the Apache-2.0 components below. Copyright remains with the original authors; this project claims no rights over them.

组件 Component来源 Source许可 License

MOSS-Image Base https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base Apache-2.0
MOSS-Image SFT https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT Apache-2.0
MOSS-Image 官方仓库 Official repo https://github.com/OpenMOSS/MOSS-Image Apache-2.0
ComfyUI 重打包 Repack https://huggingface.co/Comfy-Org/MOSS-Image Apache-2.0
文本编码器 Text encoder — Qwen3.5-9B https://huggingface.co/Qwen/Qwen3.5-9B Apache-2.0
VAE — FLUX.2-klein-base-4B https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B Apache-2.0

许可全文 / Full license text: https://www.apache.org/licenses/LICENSE-2.0

本项目同时使用 Base 与 SFT 两个官方变体，并在权重层面进行混合（见 §3.3）。
Both official variants — Base and SFT — are used, and are blended at weight level (§3.3).

2. 第三方加速权重 / Third-party acceleration weights

产物 Artifact作者 Author机制 Method许可

MOSS-Image-SFT-8-steps-dmd lodestones data-space DMD（SFT 基底）Apache-2.0
MOSS-Image-Turbo-MFDcongruency Mean Flow Distillation（基于 Base）Apache-2.0

• https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd

• https://huggingface.co/congruency/MOSS-Image-Turbo-MFD

无隶属声明 / No affiliation. 本项目与 OpenMOSS、Comfy-Org、lodestones、congruency、 toxicdog 及其关联方（含 Freepik / Magnific）无任何隶属、赞助、授权或背书关系。 No affiliation, sponsorship, authorization or endorsement by the above parties is claimed or implied.

3. 加速方案合并 / Merged acceleration strategy

本节是本版本的核心变更。项目不再采用单一加速路径，而是将三条独立路线合并为统一技术栈。

This is the substantive change in v2.0: three independent acceleration routes are merged into one stack instead of being used in isolation.

3.1 三条路线 / Three routes

路线 Route载体 Carrier加速原理 Mechanism采样要求 Sampling

• A 自训练蒸馏 LoRA（rank 128，覆盖 42 层）固定 4 步直接回归真实 latent
  Euler + simple，CFG 1，4 步 （最终版本为6-8步）

• B lodestones 8-step DMD整模型级权重data-space DMD 蒸馏Euler + simple，CFG 1，8 步

• C congruency MFD（Base）LoRA（约 72 MB）Mean Flow Distillation 4–6 步

• 未加速的官方权重仍适用其原生参数：50 步、CFG 3、sgm_uniform、shift 5 （Unaccelerated official weights keep their native recipe: 50 steps, CFG 3, sgm_uniform, shift 5.）

3.2 合并方式 / How they are combined

• 权重级融合：通过自研节点 Merge Models (INT8 ConvRot aware) 对权重线性混合， 混合前先反量化（INT8 ConvRot → BF16），混合结果为 BF16 （Weight-level blending via a custom node: linear interpolation by ratio, after dequantizing INT8 ConvRot to BF16; output is BF16.）

技术说明 / Technical note: 三条路线的训练目标不同（A 为无教师直接回归，B 为 DMD 分布匹配， C 为平均流蒸馏），因此不保证可无损叠加；任何叠加结果均属新的独立派生作品，不在原作者质量担保范围内。

The three routes use different training objectives (teacher-free regression, DMD distribution matching, mean-flow distillation) and are not guaranteed composable without loss. Any combination is a new, independent derivative outside the original authors' quality guarantees.

3.3 Base 与 SFT 混用 / Mixing Base and SFT

• 两个官方变体在同一项目内并存使用

• 支持权重级混合 W = r·W₁ + (1−r)·W₂，比例可调

• 混合只作用于二维线性投影；RMSNorm 等一维参数保留自 model1

• 混合模型会稀释加速效果，步数应相应上调至 6-NFE

Both official variants are used side by side. Weight blending is supported with a tunable ratio. Blending applies to 2-D linear projections only; 1-D parameters (RMSNorm etc.) are taken from model1. Blending dilutes acceleration, so step count must be raised accordingly.

3.4 量化与打包 / Quantization and packaging

操作 Operation说明

INT8 ConvRot 量化每层 4 个投影（qkv / o / ffn_up_gate / ffn_down）共 168 张量qkv 融合重打包proj.q/proj.k/proj.v → proj.qkv，重写量化元数据反量化INT8 ConvRot → BF16，供融合与训练使用；rope.bands 始终保留

FP32AIO 打包扩散模型 + 文本编码器 + VAE 合并为单文件 checkpoint通用布局model.diffusion_model.* / cond_stage_model.* / first_stage_model.*ComfyUI 布局model.diffusion_model.* / text_encoders.<sub>.* / vae.*

格式适用范围 / Format portability: int8_convrot、w4a8、w6a8 为 ComfyUI 专用量化布局， 其他运行时无法加载；面向通用运行时的打包仅使用 bf16 / fp16 / fp8。

int8_convrot, w4a8 and w6a8 are ComfyUI-specific layouts and are not loadable by other runtimes. Packages targeting general runtimes use bf16 / fp16 / fp8 only.

4. 修改清单（Apache-2.0 §4(b)）/ Statement of changes

相对于原始 MOSS-Image 权重，本项目做了以下实质性修改：

1. 蒸馏训练：对一个 rank-128 LoRA 执行固定 4 步的扩散蒸馏（无教师、无判别器）， 训练数据为自有数据集

2. 加速方案合并：将自训练蒸馏、lodestones 8-step DMD、congruency MFD 三条路线纳入统一技术栈（§3）

3. Base / SFT 混用：两个官方变体并存，并支持权重级按比例混合（§3.3）

4. 量化：INT8 ConvRot 量化（每层 4 个投影，共 168 张量）、FP8、W4A8、W6A8 导出

5. 布局重打包：分离 q/k/v 融合为 proj.qkv；反量化至 BF16；rope.bands 固定 FP32

6. 打包形态：AIO 单文件 checkpoint，支持通用三段式布局与 ComfyUI 布局双方案

Verbatim English summary: a rank-128 LoRA distilled for a fixed 4-step schedule on a private dataset; three acceleration routes merged into one stack; Base and SFT variants blended by ratio; INT8 ConvRot / FP8 / W4A8 / W6A8 exports; q/k/v projections fused into proj.qkv with rewritten quantization metadata; dequantization to BF16 with rope.bands pinned to FP32; AIO single-file checkpoints in both universal and ComfyUI layouts.

5. 商用状态 / Commercial status

依据 Apache License 2.0 §2 与 §3：

• ✅ 允许商业使用，包括集成进产品、SaaS、收费分发

• ✅ 允许修改与再分发

• ✅ 不要求开源衍生作品（非 copyleft）

• ✅ 附带明示专利授权

• ✅ 不限制生成内容的用途

Permitted: commercial use (including product integration, SaaS, paid distribution), modification, redistribution; no obligation to open-source derivative works; express patent grant; no restriction on the purpose of generated output.

义务 / Obligations:

1. 再分发时附带本文件与 Apache-2.0 许可全文

2. 保留原始版权与归属声明

3. 标注修改内容（§4）

4. 不得暗示原作者或第三方加速权重作者为你背书（§6）

5. 若再分发融合模型，须同时保留各上游作者的许可声明

6. 不在许可范围内 / Not covered

1. 训练数据版权 — 权重许可不延伸至训练数据本身

2. 生成内容的权利 — 肖像权、商标权、著作权由使用者自行确保合规

3. 内容合规 — 成人内容等用途受分发平台条款与当地法律约束，不受 Apache-2.0 限制亦不被其许可

4. 上游加速权重的质量 — lodestones / congruency / toxicdog 的产物均为社区或商业公司发布， 其 README 标注为进行中的工作（WIP），本项目不为其正确性、稳定性或安全性作任何担保

7. 免责声明 / Disclaimer

本衍生作品按 "AS IS" 提供，不附带任何明示或默示担保，包括但不限于适销性、特定用途 适用性与不侵权担保。作者不对因使用本作品产生的任何直接或间接损失负责。完整条款见 Apache License 2.0 §7–§8。

Provided "AS IS", without warranty of any kind, express or implied. See Apache License 2.0 §7–§8.

附录 A：v1.0 原文（保留）/ Appendix A — v1.0 as published

以下为 v1.0 声明的核心条款，原文保留以备追溯。完整版见 COMMERCIAL_NOTICE.md。

基础模型：MOSS-Image Base / SFT（Apache-2.0）、ComfyUI 重打包（Apache-2.0）、 Qwen3.5-9B 文本编码器（Apache-2.0）、FLUX.2-klein-base-4B VAE（Apache-2.0）。

修改（v1.0）：

1. 加速器蒸馏训练——rank 128 LoRA，覆盖全部 42 层的注意力、FFN、时间步与输出投影， 从纯噪声以固定 4 步调度直接回归真实图像 VAE latent（direct-distill，无教师、无判别器）， 自有数据集。

2. 三种加速策略并存——自训练蒸馏、lodestones 8-step DMD、congruency MFD（Base）。

3. Base 与 SFT 权重混用——按百分比线性插值，反量化后融合，比例可调。

4. 量化与布局重打包——INT8 ConvRot 反量化至 BF16；分离 q/k/v 融合为 proj.qkv 并重写量化元数据。

商用：Apache-2.0 允许商业使用、修改、再分发，不要求开源衍生作品，附明示专利授权， 不限制生成内容用途。义务为保留声明、标注修改、不得暗示背书。

不覆盖：训练数据版权、生成内容的权利（肖像/商标/著作权）、内容合规。

免责：AS IS，无任何明示或默示担保。

附录 B：关键技术参数 / Appendix B — technical parameters

项值DiT 结构42 层，dim 3072，24 heads，ffn 8192，128 通道 latent量化张量数168（42 层 × 4 投影）INT8 ConvRotint8_tensorwise，ConvRot groupsize 256W4A8asym_w4a8_int8，group_size 16，带 codebookW6A8w6a8_int8，group_size 16，无 codebook，打包宽度 3K/4FP8逐张量 float8_e4m3fn + weight_scale保留 FP32rope.bands训练规模rank 128

基础许可 Base license Apache License 2.0 许可全文
https://www.apache.org/licenses/LICENSE-2.0

更新日期 Updated2026-10-18上一版本 Previousv1.0（附录 A / COMMERCIAL_NOTICE.md）

如需引用本声明，请保留完整内容并注明来源。
If you reference this notice, keep it intact and cite its source.