HaDeS MOSS Turbo v1.0 BF16

- Version name: v1.0 Turbo

- File: HaDeS_MOSS_v1.0_Turbo.safetensors (BF16, 10.16 GB, SHA256 cfd7e0db798616954112604d28bcc55bf1adf56851d07f112fc786876c541e9e)

- Base model: MOSS-Image (Base)

- Type: Checkpoint (diffusion model only, no text encoder or VAE)

- Recommended settings: 8 steps, CFG 1, euler / simple, 1344 x 1792

- Images: v1.0_images/ (12 PNGs with generation data). Workflow: HaDeS_MOSS_v1.0_Turbo_core_txt2img.json (core nodes only).

## About this version

HaDeS MOSS v1.0 Turbo is an eight-step MOSS-Image checkpoint. It starts from MOSS-Image Base with two LoRAs baked in:

- HaDeS MOSS Turbo 8-steps v1.0, our published turbo (eight steps, CFG 1);

- QwenLook, our own LoRA for a cleaner, more polished finish.

No extra LoRA is needed: load it and generate in eight steps.

What is new against v0.9.4:

- Preferred in a blind test. 20 new prompts, one seed each: photographs, a chalkboard and an infographic with text, manga, pixel art, claymation, ink wash, cyberpunk, a children's book. The author chose v1.0 in 16 of 20 pairs.

- Less appearance bias. On "photograph of a woman sitting in a café" with no ethnicity, 16 seeds gave about 1 East Asian face in 16, against 9-11 for v0.9.4. The women still look much alike (young, long wavy brown hair, smiling), so describe the person when it matters.

- Sharper. The QwenLook LoRA roughly doubles the measured fine detail, on photographs and on illustrations. The finish is cleaner and more polished. For a rougher, more documentary look, describe the camera and the light in the prompt.

Tip: long, detailed prompts suit it well: where the light comes from, the lens, the setting, small imperfections. The sample images use prompts written with the detailed mode of the HaDeS MOSS Suite's Prompt Enhancer (free, in ComfyUI). Its random mode writes a new prompt for each seed from an empty box.

int8 version

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:

• Same compositions in all 10 pairs. Mean absolute difference on a 64 px thumbnail: 2.9 % (1.3-4.2 %), smaller than the 4.7 % between the baked file and the live LoRAs above. Small details change: a shepherd gains a staff, a figure moves, invented labels read differently.

• Text held in both: "OPEN" and "VISIT THE FJORDS" were spelled the same.

• Sharpness about the same: mean Laplacian variance 862 against 918 for BF16, some images higher and some lower.

• A seed gives the same image again on the same setup (re-rendered pixel-identical).

Speed depends on the ComfyUI version and memory setup and was not measured as a benchmark for this card.

How to use it:

- Files:

- put the checkpoint in models/diffusion_models/;

- add the Qwen3.5-9B text encoder qwen3.5_9b_int8_convrot, Load CLIP type moss_image) and flux2-vae from Comfy-Org/MOSS-Image.

- Stock nodes: Load Diffusion Model, then KSampler with 8 steps, CFG 1, euler / simple. Do not add a turbo LoRA: it is already inside.

- HaDeS MOSS Suite: MOSS Loader and MOSS KSampler, with the turbo input left empty.

Not measured yet: other step counts, CFG above 1, other sizes, the int8 version. The author's impression is that it looks best above about 1.5 MP and that CFG 1.2-1.5 can suit it.

Sample images: the author's renders at 1344 x 1792 and seed 1001, with the settings above. They were made with the release candidate, which has the same weights; a re-render with the released file was pixel-identical, so the generation data names the released file.

Full details: [model card on Hugging Face](https://huggingface.co/DeVilDoNia/HaDeS-MOSS-v1.0-Turbo) · [measurement log](https://github.com/Devildonia/HaDeS-MOSS-Suite/blob/main/tools/RESULTS.md) · [HaDeS MOSS Suite](https://github.com/Devildonia/HaDeS-MOSS-Suite) · [Turbo LoRA](https://huggingface.co/DeVilDoNia/HaDeS-MOSS-Turbo)

License: Apache 2.0. Built on MOSS-Image Base by the OpenMOSS Team (Apache 2.0); the QwenLook training images were made with Qwen-Image-Edit 2509 and Qwen-Image-Lightning (Apache 2.0).

HaDeS MOSS Turbo v0.9.4

A single-file, eight-step MOSS-Image text-to-image checkpoint for ComfyUI. It starts from MOSS-Image-Base, with three LoRAs baked into the diffusion weights. It offers a different rendering choice from MOSS-Image-SFT; the comparisons below do not show that either model is better for every prompt. No separate Turbo or other LoRA is needed to run this file.

This is only the diffusion model. It does not contain the MOSS text encoder or VAE.

Use in ComfyUI

1. Put the .safetensors file in ComfyUI/models/diffusion_models/ and select it in a MOSS-compatible loader. The measured workflow used the HaDeS MOSS Suite MOSS Loader.

2. Load the MOSS-compatible qwen3.5_9b_int8_convrot.safetensors text encoder and flux2-vae.safetensors VAE. Both are available in Comfy-Org/MOSS-Image. These are the exact companion files used in the comparisons here.

3. Enter a text-to-image prompt, create an empty latent, sample, then decode with the VAE. If using the Suite, set MOSS Prompt Enhancer to as written or caption, connect MOSS Latent and MOSS KSampler, and use these starting settings:

Setting Measured reference

Size: 1/2 Mpx

Steps: 8

CFG: 1.0

Sampler: Euler

Scheduler: Simple

Flow shift: 5.0 (0 in HaDeS MOSS Suite)

Denoise 1.0

SLG: Off

Leave the KSampler's separate turbo LoRA input unconnected: its Turbo weights are already in this checkpoint. These settings were exercised on ComfyUI 0.43.0. Other resolutions works perfectly, but the comparisons used 1152 × 1536.

CFG 1 is the reproducible starting point, not a universal aesthetic optimum. In a paired four-scene, two-seed screen, raising CFG from 1 to 2 at eight steps increased saturation and near-clipped pixels, darkened all eight images, and took about 1.8 times as long in that session. Some reviewers preferred stronger CFG on particular images. Increasing steps also changed compositions; it was not a consistent quality improvement.

Other step counts and CFG

Measured on eight scenes, one seed each (measurement log): no image broke or burned from 6 to 50 steps at CFG 1, nor up to CFG 3 at 20 steps. More steps add texture and fine detail and take longer (about 4 s at 8 steps, 9.5 s at 20 and 23 s at 50 on an RTX 5090 at 1152 × 1536), but the image does not settle: each step count gives a different image of the same scene, so pick the count you like rather than expecting more steps to refine the 8-step one. In one scene a short sign was only right from 12 steps. Higher CFG darkens and saturates the image and doubles the time per step.

Construction

Check the model card on Hugging Face for more details: https://huggingface.co/DeVilDoNia/HaDeS-MOSS-v0.9.4-Turbo

Some Quick Start Workflows

https://civitai.com/models/3039368/hades-moss-suite-workflows