Newer version: HaDeS MOSS v1.0 Turbo (Base + Turbo v1.0, about 1 East Asian face in 16 on the neutral café prompt against 9-11 here). This file stays available.
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.
| Property | Value |
|---|---|
| File | HaDeS_MOSS_v0.9.4_Turbo.safetensors |
| Format | ComfyUI MOSS-Image diffusion model, single-file safetensors |
| Precision | BF16; 10.16 GB (9.47 GiB) |
| SHA-256 | 32c215fc08ee8ec0e55459507e0712936adccf5804e94b6a7264bc33efd68770 |
| Base checkpoint | MOSS-Image-Base, not MOSS-Image-SFT |
| Author | HaDeS |
This is only the diffusion model. It does not contain the MOSS text encoder or VAE.
These are display-size JPEGs made from BF16 v0.9.4 outputs in the eight-scene evaluation, not images from the SFT comparison arm. All six originals used the settings below at 1152 × 1536; the previews are resized to 768 × 1024. Seeds and source/preview checksums are in assets/manifest.json. These examples show individual outcomes, not a measured success rate.
| Watchmaker · seed 2401 | Climber · seed 2403 | Flower seller · seed 2404 |
|---|---|---|
| Rainy alley · seed 2406 | Citrus still life · seed 2407 | Lunar observatory · seed 2408 |
|---|---|---|
.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.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.MOSS Prompt Enhancer to as written, connect MOSS Latent and MOSS KSampler, and use these starting settings:| Setting | Measured reference |
|---|---|
| Size | 1152 × 1536 |
| Steps | 8 |
| CFG | 1.0 |
| Sampler / scheduler | Euler / simple |
| Flow shift | 5 |
| Denoise | 1.0 |
| Skip-layer guidance | 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 may work, but the comparisons below 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. The author preferred stronger CFG on some images (8 / CFG 2 on 3 of 8 sheets). Increasing steps also changed compositions; it was not a consistent quality improvement.
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.
Asked for "a woman sitting in a café" with no ethnicity, 16 seeds at 8 steps gave about 9-11 of 16 East Asian faces (counted by eye), against about 2-4 for MOSS-Image-Base at 25 steps. The Turbo LoRA baked into this model carries that bias. Name the person's appearance in the prompt when it matters: MOSS-Image-SFT and Base followed an explicit ethnicity in every test of the measurement log (not yet checked on this model).
The author's saved workflow has this active path: MOSS-Image-Base BF16 → MOSS Multi-LoRA → MOSS Model Saver. The LoRAs were applied in this order:
| LoRA | Strength | Notes |
|---|---|---|
data_expanded.safetensors |
0.30 | Expanded-dataset LoRA described by the author as trained with MOSS-Image-SFT, then applied here to Base. |
HaDeS_MOSS_color_detailer_v0.2.safetensors |
0.60 | A merged color/detail LoRA. Its file metadata names HaDeS_test_colour_v01.safetensors and moss_context_detaile_tweaker_20261018_rank2.safetensors, combined at inner weights 0.6 and 0.75. The table's 0.60 is the separate outer application strength. |
HaDeS-MOSS-BeTa-Turbo-8steps-v0.9.safetensors |
1.00 | The author's unreleased Turbo v0.9 LoRA, baked into the checkpoint. The LoRA header says steps: 4, while this checkpoint was evaluated at eight inference steps. |
The model was saved as BF16 with nearest rounding. Its safetensors header identifies MOSS-Image architecture and 42 layers, but does not embed the recipe; this section records the author's workflow and the component-file metadata. The standalone LoRAs are not required for inference. No claim is made here that the recipe alone will reproduce a bit-identical file from different software versions or source files.
The measurement log records exact prompts, seeds, settings, image metrics and the author's blind votes. Comparisons used identical prompt text, seed, 1152 × 1536 size, 8 steps, CFG 1, Euler/simple, flow shift 5, text encoder and VAE within each image pair. The compared pipelines differed in their model weights and LoRA application; these tests do not isolate the contribution of any one baked LoRA.
| Comparison | Images | Author's blind preference for v0.9.4 |
|---|---|---|
| Against SFT + private Turbo v0.9, eight new prompts fixed before rendering | 8 pairs | 5/8 |
| Against SFT + released Turbo v0.7, same eight prompts and seeds | 8 pairs | 6/8 |
One reviewer, one seed per prompt; the second comparison reused familiar images from the first, so it was only partially blind. AI-model reviews of the same sheets are logged separately as qualitative impressions and are not counted here. The author found v0.9.4 publishable as an alternative to SFT, and the results support that practical description. They do not establish general superiority, long-run consistency or a model-wide defect rate. In the eight new pairs, v0.9.4 had higher measured mean HSV saturation in every pair and more near-clipped pixels in seven; those numbers describe appearance, not image quality. Hands, small text and physical structure still deserve inspection in individual outputs.
The comparisons used this BF16 file. They do not validate the visual equivalence of the separately saved int8 ConvRot variant. The custom LoRA datasets and training details are not fully specified by the available model metadata; do not infer a training-data inventory from this card.
Built on OpenMOSS-Team/MOSS-Image-Base, whose model weights are released under Apache 2.0. This HaDeS checkpoint is also released under Apache 2.0. See the MOSS-Image technical report and the HaDeS MOSS Suite for the upstream model and ComfyUI tooling.