文件清单 线路:原始(huggingface.co) · 切换
下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。
说明
Moss_image_颜汐
作者:YI_MIAN_668
YanXi · Spectrum Recast
moss_image_YanXi_ssx_fp16 — a full-SSX fusion bake on MOSS-Image-SFT.
Not a LoRA you load at runtime. A recipe melted into the weight spectrum itself. 164 projection layers dequantized one by one, three incompatible LoRA formats normalized onto the same tensors, then a full-spectrum SVD tail expansion on every single one. Spectral coverage 53.57% → 94.35%. Energy gain positive on all 164 layers, zero exceptions.
01 / What this is
An offline, weight-level fusion bake — not a runtime LoRA stack, not a naive weight sum. The base is a tensor-wise int8 + Hadamard convrot quantized MOSS-Image-SFT. We dequantize it out of the codebook into float, normalize three mutually incompatible recipes onto the same projection tensors (a LoKr Kronecker product, a fused-qkv low-rank product, and q/k/v three-segment row slices), then give every fused layer a full-spectrum SVD with tail-spectrum energy redistribution before writing it back in FP16. All 164 fused projection layers, 100% expanded, zero skipped (ALL_LAYERS). Layers outside the fusion set are copied byte-for-byte at original precision — the other half of the fidelity story.
02 / The bake, as a formula
W* = FP16_cast( D_convrot(Q_base, scale) + S_SSX( SUM_l alpha_l * Pi_l(A_l, B_l) ) )
S_SSX(D) = U diag( sigma_head , max(sigma_tail, tau) lambda ) * V^T
r = min k s.t. SUM_{i<=k} sigma_i^2 >= 0.99 * E (E = ||D||^2_F)
tau = max( 1.5 sigma_r , sqrt( 0.08 E / (m - r) ) )
lambda = 0.20 * E / (E' - E)
C_spec = r_after / min(N, K) -> 53.57% -> 94.35%
Pi_l is the per-format operator: W1 (x) W2 for LoKr, B A for fused qkv, and narrow(0, i*N/3, N/3) row-slice accumulation for q/k/v splits. EFF=0.99, BETA=1.5, TAIL=0.08, BUDGET=0.20. Full-spectrum torch.linalg.svd(full_matrices=False), never a truncated svd_lowrank. In one sentence: decompose the recipe's delta across the full spectrum, give the directions that should have stayed quiet a share of the energy back, then clamp the total with a 20% energy budget. Reinforcement, not amplification — and specifically not flattening the spectrum into noise.
03 / Verified numbers (164 per-layer records)
- Fused / SSX layers: 164 / 164, 100% coverage, zero skipped
- Spectral coverage C_spec: 53.57% → 94.35% (per-layer 91.37%–96.81%)
- Effective rank / 3072: mean 1645.6 → 2898.3 (per-layer 627–2337 → 2807–2974)
- Per-layer energy gain: +7.00% to +14.77%, mean +7.84%, all 164 layers > 0
- FP16 fidelity: relL2 = 2.078e-4 (~99.98%), delta direction cos 0.99918–0.99980
- Recipe perturbation: ||delta|| / ||W_base|| = 0.37%–1.08% — gentle, not an overwrite
- q/k/v row slices: 41 qkv layers x 3 segments, 41 fusion events, 9216 rows fully covered
- Numerical anomalies: NaN / Inf layers = 0
- Artifact: 429 keys, 9.361 GiB, F16x164 / I8x4 / BF16x210 / F32x5 / U8x46, full run 501 s Triple-checked: offline numerical pre-screen gate, 4-layer pilot, full-run sampled recomputation. The recomputation error matches the FP16 rounding floor exactly.
04 / What it looks like
Built for painterly output. Anime figures, stacked gouache and oil pigment with visible impasto and granulation, and bloom / rim-light / transmitted glow are consistently strengthened. The overall feel leans toward the high-saturation, airy, edge-lit illustration look of the niji / krea2 family — vivid color that stays translucent rather than muddy. Not a sharpening or contrast trick applied afterwards: it comes from reallocating energy across the weight spectrum, so low-energy directions participate again and fine texture plus glow detail resolve more completely. Please judge the final look by same-seed A/B and your own eyes.
05 / Recommended settings (MOSS-Image-SFT)
Model: moss_image_YanXi_ssx_fp16.safetensors (no LoRA loaded, no trigger word)
Sampler: euler Scheduler: sgm_uniform Flow shift: 5.0
Steps: 50 CFG: 3.0
VAE: flux2-vae Text enc.: Qwen3.5-9B (stock)
Resolution: ~1 MP, multiples of 16 (1024x1024 / 848x1200 / 1200x848)
Negative: worst quality, bad anatomy, watermark, jpeg artifact, nsfw
MOSS-Image uses a Qwen3.5-9B LLM as text encoder: write long descriptive sentences (subject → environment → lighting → style), not tag soup, not "masterpiece, best quality" boilerplate, and no (word:1.5) weight syntax. Text you want rendered goes in quotes. Put the file in models/unet/ and load it directly in ComfyUI. low-VRAM mode;and add --fp16-unet if the weights get cast to BF16.
06 / Credit — this is not a solo work
Made by an AI working team in collaboration with the publisher. It is not the work of one person working alone.
- AI working team: int8_convrot dequantization, normalization of the three heterogeneous key formats, the full-spectrum SVD tail-expansion operator and its energy budget, the offline numerical gate, the 164-layer full bake, and all verification and fidelity checks.
- Publisher: recipe selection, weight decisions, iteration trade-offs, and the final call on how the images actually look. Methodology belongs to the team; aesthetic judgment belongs to the human. engineering collaboration, not a hand-made solo release. The file's metadata carries model_name / modelspec.title plus the original bake_flow, bake_loras, bake_ssx and bake_qkv_split records, so the build is traceable and reproducible. ---
YanXi · 谱域重铸(中文 / Chinese)
moss_image_YanXi_ssx_fp16 —— 基于 MOSS-Image-SFT 的全量 SSX 融合烘焙。
不是把 LoRA 挂在模型上,而是把配方熔进权重的谱里。 164 个投影层逐层反解、三路互不兼容的 LoRA 格式归一到同一批张量, 再对每一层做全谱 SVD 尾谱扩展。 谱能量覆盖率 53.57% → 94.35%,164 层逐层能量增益全部为正,无一例外。
01 / 这是什么
一次离线权重级融合烘焙——不是运行时挂载,也不是把几个 LoRA 的权重简单相加。 基座是张量级 int8 + Hadamard 卷积旋转(int8_convrot)量化的 MOSS-Image-SFT。 我们先把它从量化码本反解回浮点域,再把三路结构互不兼容的配方逐层归一到同一批投影张量: LoKr 的 Kronecker 积、融合 qkv 的低秩积、q/k/v 三段行切片映射; 随后对每个融合层做全谱奇异值分解与尾谱能量重分配,最终以 FP16 落盘。 全部 164 个参与融合的投影层,100% 执行尾谱扩展,零跳过(ALL_LAYERS)。 未参与融合的层按原始精度逐字节保留,基座结构零扰动——这是保真的另一半。
02 / 烘焙公式
W* = FP16_cast( D_convrot(Q_base, scale) + S_SSX( SUM_l alpha_l * Pi_l(A_l, B_l) ) )
S_SSX(D) = U diag( sigma_head , max(sigma_tail, tau) lambda ) * V^T
r = min k s.t. SUM_{i<=k} sigma_i^2 >= 0.99 * E (E = ||D||^2_F)
tau = max( 1.5 sigma_r , sqrt( 0.08 E / (m - r) ) )
lambda = 0.20 * E / (E' - E)
C_spec = r_after / min(N, K) -> 53.57% -> 94.35%
Pi_l 是各路键格式对应的算子:LoKr 用 W1 (x) W2,融合 qkv 用 B A, q/k/v 拆分按 narrow(0, i*N/3, N/3) 行切片累加。 参数 EFF=0.99 / BETA=1.5 / TAIL=0.08 / BUDGET=0.20; 使用全谱 torch.linalg.svd(full_matrices=False),绝不用截断的 svd_lowrank。 一句话:把配方增量在全谱上分解,给本该沉寂的方向重新分到能量, 再用 20% 的能量预算锁死总强度。是「强化」,不是「放大」,更不是把谱摊平成白噪声。
03 / 验收数据(164 行逐层记录)
- 融合层 / SSX 层:164 / 164,100% 覆盖,零跳过
- 谱能量覆盖率 C_spec:53.57% → 94.35%(逐层 91.37%~96.81%)
- 有效秩 / 3072:均值 1645.6 → 2898.3(逐层 627~2337 → 2807~2974)
- 逐层能量增益:+7.00% ~ +14.77%,均值 +7.84%,164 层全部 > 0
- FP16 保真:relL2 = 2.078e-4(约 99.98%),增量方向 cos 0.99918~0.99980
- 配方扰动:||delta|| / ||W_base|| = 0.37%~1.08%,温和非暴力覆盖
- q/k/v 行切片:41 个 qkv 层 × 3 段,41 次融合事件,9216 行全覆盖
- 数值异常:NaN / Inf 层 = 0
- 产物:429 键、9.361 GiB,F16×164 / I8×4 / BF16×210 / F32×5 / U8×46,全量耗时 501 s 经三重校验:离线数值预筛闸门、4 层试点、全量抽层重算; 重算误差与 FP16 舍入量级精确吻合。
04 / 画面取向
以绘画性为轴心:二次元造型、颜料堆叠的厚涂笔触与颗粒质感、 以及辉光 / 透光 / 边缘泛光的光晕层次被显著强化; 整体趋近 krea2 的 niji 系那种高饱和、色彩通透、轮廓带光的插画观感—— 浓烈但依旧通透,不发闷。 这不是锐化或提对比堆出来的后期效果,而是权重谱域重分配后, 低能量方向重新参与响应带来的纹理与光感完整度。 具体观感请以同 seed A/B 与人工目视为准。
05 / 推荐参数(MOSS-Image-SFT)
Model: moss_image_YanXi_ssx_fp16.safetensors (no LoRA loaded, no trigger word)
Sampler: euler Scheduler: sgm_uniform Flow shift: 5.0
Steps: 50 CFG: 3.0
VAE: flux2-vae Text enc.: Qwen3.5-9B (stock)
Resolution: ~1 MP, multiples of 16 (1024x1024 / 848x1200 / 1200x848)
Negative: worst quality, bad anatomy, watermark, jpeg artifact, nsfw
MOSS-Image 用 Qwen3.5-9B 大语言模型作文本编码器:写完整的描述性长句 (主体 → 环境 → 光线 → 风格),不要 tag 汤,不要「masterpiece, best quality」这类咒语, 也不要 (word:1.5) 权重语法;要渲染在画面里的文字用引号包住。 文件放进 models/unet/,ComfyUI 直接加载。 低显存模式:若权重被转成 BF16,启动加 --fp16-unet。
06 / 署名 — 这不是个人作品
本模型由 AI 工作团队 × 本模型发布者 协作完成,并非由个人独立制作。
- AI 工作团队:int8_convrot 量化域反解、三路异构键格式归一、全谱 SVD 尾谱扩展算子与 能量预算约束、离线数值预筛闸门、164 层全量烘焙,以及全部结构与保真校验。
- 发布者:配方选型、权重决策、多轮迭代取舍、画面观感的最终裁定。 方法论归团队,审美判断归人。这是一次人机协同的工程化实验。 文件的 metadata 里写有 model_name / modelspec.title, 以及原始的 bake_flow、bake_loras、bake_ssx、bake_qkv_split 记录,可追溯、可复现。
