A single-file, eight-step MOSS-Image text-to-image checkpoint for ComfyUI. It starts from MOSS-Image-Base with two LoRAs baked into the diffusion weights: the published HaDeS MOSS Turbo 8-steps v1.0 and the author's QwenLook LoRA. No separate Turbo or other LoRA is needed to run this file. It follows HaDeS MOSS v0.9.4 Turbo, which stays available.
| File | Precision | Size | SHA-256 |
|---|---|---|---|
HaDeS_MOSS_v1.0_Turbo.safetensors |
BF16 | 10.16 GB (9.47 GiB) | cfd7e0db798616954112604d28bcc55bf1adf56851d07f112fc786876c541e9e |
HaDeS_MOSS_v1.0_Turbo_int8_convrot.safetensors |
int8 ConvRot | 5.41 GB (5.04 GiB) | e0707a518cb8b81bf435133b45808bfd0a849851d12689ad8021748c25ca7344 |
Both are ComfyUI MOSS-Image diffusion models in single-file safetensors, built on MOSS-Image-Base (not SFT), by HaDeS. The int8 file is the BF16 one converted, with no other change (see int8 version). These are only the diffusion model; they do not contain the MOSS text encoder or VAE.
Display-size JPEGs (768 × 1024) of the author's outputs at 1344 × 1792, seed 1001, with the settings below. The prompts are long, detailed descriptions written with the detailed mode of the HaDeS MOSS Suite Prompt Enhancer; the full prompts, settings and checksums are in assets/manifest.json. They were rendered with the release candidate of this file, which has the same weights (a re-render with the released file was pixel-identical). These examples show individual outcomes, not a measured success rate.
| Diner at blue hour | Elder in the desert | Shepherd at dawn |
|---|---|---|
| Travel poster | Anime key visual | Flamenco |
|---|---|---|
.safetensors file in ComfyUI/models/diffusion_models/ and load it with Load Diffusion Model (or the Suite's MOSS Loader).qwen3.5_9b_int8_convrot.safetensors text encoder (Load CLIP, type moss_image) and the flux2-vae.safetensors VAE from Comfy-Org/MOSS-Image. These are the companion files used in every test below.| Setting | Tested reference |
|---|---|
| Size | 1344 × 1792 |
| Steps | 8 |
| CFG | 1.0 |
| Sampler / scheduler | Euler / simple |
| Flow shift | 5 |
| Skip-layer guidance | Off |
Leave any separate Turbo LoRA off: it is already in this checkpoint. Every test on this card used these settings. Not measured on this model: other step counts, CFG above 1, other sizes. The author's impressions, not measured: it looks best above about 1.5 MP, and CFG 1.2-1.5 can suit it.
MOSS-Image-Base BF16 → LoRAs below → MOSS Model Saver (BF16, stochastic rounding). The recipe is written in the file header (hades/recipe).
| LoRA | Strength | Notes |
|---|---|---|
| HaDeS MOSS Turbo 8-steps v1.0 | 1.00 | The author's published eight-step DMD LoRA, trained on Base. |
| QwenLook Base | 1.00 | The author's text-to-image LoRA, not published separately: rank 32, 1000 steps at 1024 px, on 1833 images made by Qwen-Image-Edit 2509 (with the Qwen-Image-Lightning LoRA) from MOSS-Image-SFT scenes, each captioned with its scene prompt. |
Against the same two LoRAs applied live in ComfyUI (20 prompts, same seeds), the baked file gives the same compositions and the same sharpness (mean Laplacian variance 77.7 against 77.1); small details can differ, from rounding amplified by the eight steps (in one image a pistol in a hand was gone).
HaDeS_MOSS_v1.0_Turbo_int8_convrot.safetensors is the BF16 file saved by the Suite's MOSS Model Saver in the layout of Comfy-Org's *_int8_convrot files: the linear layers of the 42 blocks in int8 with ConvRot (Hadamard rotation) and int8 attention, everything else in BF16. It has the same keys, dtypes and shapes as Comfy-Org's moss_image_sft_int8_convrot. Load it the same way, with the same settings; it needs a ComfyUI recent enough to load those files.
Against the BF16 file, 10 prompts (photographs, an oil painting, anime, a manga page, a 3D animated still, a travel poster with text, a scientific plate), seed 1001, 1344 × 1792, 8 steps, CFG 1:
Speed depends on the ComfyUI version and memory setup and was not measured as a benchmark for this card.
Against v0.9.4 Turbo: 20 new prompts (photographs, text on a chalkboard and an infographic, manga, pixel art, claymation, ink wash, cyberpunk, a children's book), one seed each, 8 steps, CFG 1, 1344 × 1792, shown in pairs with left / right at random. The author preferred v1.0 in 16 of 20 pairs and v0.9.4 in 4. One reviewer, one seed per prompt: this shows a preference on these prompts, not general superiority. An AI review of the same pairs is logged separately and not counted here.
What the QwenLook LoRA changes (Base + Turbo v1.0 with and without it, 16 photograph and 4 illustration prompts, one seed each):
Known bias: asked for "photograph of a woman sitting in a café" with no ethnicity, 16 seeds at 8 steps gave about 1 East Asian face in 16 (counted by eye), against about 9-11 for v0.9.4. The 16 women look much alike, though: young, long wavy brown hair, smiling at the camera. Describe the person when it matters.
Hands, small text and crowded scenes still deserve a look in individual outputs.
Built on OpenMOSS-Team/MOSS-Image-Base, released under Apache 2.0. The QwenLook training images were made with Qwen/Qwen-Image-Edit-2509 and lightx2v/Qwen-Image-Lightning, both Apache 2.0. This checkpoint is released under Apache 2.0. ComfyUI tooling: HaDeS MOSS Suite.
HaDeS_MOSS_v1.0_Turbo.safetensors 10 GBHaDeS_MOSS_v1.0_Turbo_int8_convrot.safetensors 5.4 GBOpenMOSS-Team/MOSS-Image-Base