MOSSKVA — MOSS-Image SFT Merge (INT8 ConvRot). ADVANCED HANDCRAFTED MERGE based on MOSS-Image SFT (5B DiT — Diffusion Transformer with Flow Matching), blending three specialized LoRAs for anatomy and raw amateur realism. MERGE RECIPE (calibrated to avoid overbaking): MysticXXX Moss V1 (NSFW & Anatomy) 0.60 | Lenovo Moss (Primary Amateur Photorealism) 0.50 | NiceGirls V2 Moss (Secondary Photorealism) 0.30. FEATURES: Hyper-realistic skin texture eliminating the plastic 3D render look — Natural amateur lighting: smartphone flash, low light, authentic colors — Uncensored anatomical understanding without forcing explicit content — Keeps casual dressed portrait capability. QUANTIZATION (INT8 ConvRot — 5.04 GB): TensorWiseINT8Layout with regular Hadamard ConvRot rotation (group 256) — Fidelity: practically identical to the original BF16 (10.05 GB), CosSim = 0.9933 — Direct GPU quantization using native ComfyUI/comfy-kitchen layouts. REQUIRED FILES IN COMFYUI: Text Encoder (models/text_encoders/): qwen3.5_9b_w4a8_gptq.safetensors (CLIPLoader, type: moss_image) — VAE (models/vae/): flux2-vae.safetensors (VAELoader to VAEDecode) — Empty Latent: EmptyFlux2LatentImage (channels=128, downscale=16x). RECOMMENDED SETTINGS (Standard SFT): Steps 30-50 (ideal 35) — CFG Scale 3.0-4.0 (default 3.0) — Sampler: euler or dpmpp_2m_sde — Scheduler: sgm_uniform or beta — Resolution 1024x1024 / 896x1152 — Model Sampling Shift 5.0 (AuraFlow / Flow) — Average time on RTX 5060 Ti: 8-15 sec per image. PROMPT TRIGGERS FOR AMATEUR REALISM: "candid amateur phone photo, casual selfie taken with an iPhone/Samsung" — "dim room lighting, direct flash photography, authentic harsh shadows" — "natural uneven skin texture, subtle skin imperfections, visible pores, no makeup" — "spontaneous pose, slightly grainy film look, messy background, candid snapshot". AVOID in positive prompt: masterpiece, 8k, hyperrealistic, octane render, flawless skin, cinematic lighting (forces artificial smooth video-game skin). VERSIONS: mosskva 1.0 (INT8 ConvRot, this version, 5.04 GB) — mosskva 1.0 (W6A8 ConvRot, 4.21 GB) — mosskva 1.0 (NVFP4 Hybrid, 4.20 GB). Additional Turbo DMD2 versions (8 steps, CFG 1.0) also available in this series. LICENSE: allows use without credit, allows sharing merges, allows different licenses on merges. Commercial use allowed. Not intended to depict actual persons.
Description
WHAT THIS VERSION IS:
mosskva w6a8 turbo dmd2 is a W6A8 quantized variant (4.21 GB) of the mosskva 1.0 Turbo DMD2 merge, built on MOSS-Image SFT (5B DiT with Flow Matching). It uses AsymW4A8Int8Layout in 6-bit mode (w6a8_int8) with 256-channel Hadamard ConvRot rotation, per-group MSE-optimal scale search (group size 16), and INT8 activations. Cosine similarity to the INT8 Turbo DMD2 reference is 0.9829 — virtually indistinguishable while saving ~850 MB VRAM. This version is DISTILLED via DMD2 (8-step) for ultra-fast generation at CFG 1.0.
HOW TO LOAD IN COMFYUI:
- Loader: UNETLoader (standard) — loads the .safetensors directly as a checkpoint
- Text Encoder: qwen3.5_9b_w4a8_gptq.safetensors (models/text_encoders/) via CLIPLoader with type="moss_image"
- VAE: flux2-vae.safetensors (models/vae/) via VAELoader
- Latent: EmptyFlux2LatentImage (channels=128, downscale=16x)
MERGE RECIPE (fused in Float32 over BF16 base, then GPU-quantized to W6A8 ConvRot):
- MysticXXX Moss V1 (NSFW & Anatomy): strength 0.60
- Lenovo Moss (Amateur Photorealism): strength 0.50
- NiceGirls V2 Moss (Secondary Photorealism): strength 0.30
- 8-Steps DMD2 LoRA (Flow distillation): strength 1.00
TURBO DMD2 SPECIFIC SETTINGS (8 steps, CFG 1.0 — negative prompt ignored):
- Steps: 6-10 (ideal 8)
- CFG: 1.0 (MANDATORY — negative prompt is NOT processed)
- Sampler: euler or dpmpp_sde
- Scheduler: sgm_uniform or simple
- Resolution: 1024x1024 / 896x1152
- ModelSampling Shift: 3.0-5.0
- Avg time: ~2-3s/image (RTX 5060 Ti)
ADVANTAGES VS INT8 TURBO DMD2 (5.04 GB):
- ~850 MB smaller (4.21 GB vs 5.04 GB) — fits in tighter VRAM budgets
- CosSim 0.9829 — visual quality virtually identical to INT8 Turbo
- Faster load times, lower bandwidth
TRADE-OFFS / WHEN INT8 TURBO IS STILL BETTER:
- INT8 ConvRot Turbo (CosSim 0.9933) remains the fidelity reference
- Extreme edge-case texture fidelity may marginally favor INT8
- If VRAM is not a constraint, INT8 Turbo is the safe default
TRIGGER WORDS (recommended):
candid amateur phone photo, smartphone photo, direct flash, authentic harsh shadows, natural uneven skin texture, subtle skin imperfections, visible pores, no makeup, spontaneous pose, slightly grainy, messy background, candid snapshot
IMPORTANT: Do NOT use words like "masterpiece, 8k, hyperrealistic, octane render, flawless skin, cinematic lighting" — these force artificial smooth skin and defeat the amateur realism aesthetic.
TECHNICAL DETAILS:
- Base Model: MOSS Image
- Precision: W6A8 (6-bit weights, 8-bit activations) with ConvRot (Hadamard 256)
- Quantization: AsymW4A8Int8Layout (w6a8_int8), per-group MSE scale search (group 16)
- File: moss_image_sft_turbo_dmd2_mystic_lenovo_nicegirls_w6a8.safetensors (4.21 GB)
- Requires: qwen3.5_9b_w4a8_gptq.safetensors (text encoder), flux2-vae.safetensors (VAE)