HaDeS MOSS Turbo, version v1.0

• Version name: Turbo 8-steps v1.0

• File: HaDeS-MOSS-Turbo-8steps-v1.0.safetensors (3.56 MB, SHA256 fccd4ef0debca8f0da8b111224cfbcd14b1ae835d7b55996841f86be08872f02)

• Base model: MOSS-Image (Base)

• Recommended settings: 8 steps, CFG 1 (or 1.5), euler / simple, strength 1.0

About this version

Turbo 8-steps v1.0 is an 8-step LoRA for MOSS-Image Base. It is our best turbo so far. It is a 3.5 MB file: rank 2, on layers 24-41 only. It was trained with the HaDeS MOSS Suite's own scripts, with no image dataset.

What is new: less appearance bias. Our earlier turbos drifted toward East Asian faces on prompts that do not describe the person. v1.0 was trained to keep more of Base's variety. Counted on "photograph of a woman sitting in a café", 16 seeds:

• MOSS-Image Base at 25 steps: about 2-4 of 16 East Asian faces;

• our previous 8-step turbo: about 11-12;

• v1.0: about 8.

The bias is lower, but it is not gone. Describe the person in the prompt when you want a specific look; the model follows it.

Two looks from one LoRA:

• CFG 1: light and soft, with less saturation than our previous turbo.

• CFG 1.5: more contrast, colour and sharpness. The scene may change a little, because the guidance reads the prompt more literally. Each step runs the model twice.

The negative prompt changes little. At CFG 1 ComfyUI ignores it. At CFG 1.5, an empty negative and a typical one gave nearly the same image.

Text: on our test set, the sign, the three-line poster and the shop signs in the scenes were all right, at CFG 1 and at CFG 1.5.

How to use it:

• Model: moss_image_base_bf16 from Comfy-Org/MOSS-Image, with the Qwen3.5-9B text encoder and the FLUX.2 VAE.

• Stock nodes: LoraLoaderModelOnly, then KSampler with 8 steps, CFG 1 or 1.5, euler / simple, and strength 1.0.

• HaDeS MOSS Suite: a MOSS Turbo LoRA node into the turbo input of MOSS KSampler, with its steps set to 8 (and its CFG to 1.5 for the second look).

Also on SFT: I used it on MOSS-Image SFT with good results (see the Base / SFT sheet); not measured.

Not measured yet: other strengths or step counts, sizes above 1536 px, img2img and editing.

Full measurements: [model card on Hugging Face](https://huggingface.co/DeVilDoNia/HaDeS-MOSS-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)

License: Apache 2.0. Trained from MOSS-Image Base by the OpenMOSS Team (Apache 2.0).

Few-step MOSS-Image in 3.4 MB.

Three few-step LoRAs for MOSS-Image SFT that we trained from zero on one RTX 5090, plus 3.6 MB cuts of two community LoRAs, credited below. All files are rank 2 on layers 24–41 only.

Ours (HaDeS):

• 4steps - v0.6: the most robust on text at 4 steps (layout JSON 6 / 6, a 16:9 three-line poster 3 / 3 in our tests).

• 4steps - v0.6 Diverse: keeps more of SFT's own variety on open prompts (faces, poses, props).

• 8steps - v0.7: for 8 steps, up to 1024 px. Judged blind in pairs at 8 steps on 20 varied prompts, it beat akhaliq's 4-step LoRA (full 1.2 GB file) 18 to 5, and matched lodestones' Base / Kromoss 8-step LoRA run on SFT at 1024 (10 to 7). At 2048 that lodestones LoRA wins (5 to 0).

Not ours (credited cuts, Apache 2.0):

• akhaliq 4-step v2.0 (cut): akhaliq's MOSS-Image-SFT-4-steps v2.0, reduced to layers 24–41 at rank 2.

• lodestones Kromoss 8-step DMD2 (cut): lodestones' 8-step DMD2 LoRA distilled from MOSS-Image Base (shipped with his Kromoss finetune), reduced the same way. It is a Base / Kromoss LoRA; on SFT it runs, and it is what we compared against.

The cuts render the original files' images (by eye, on our test prompts; colour within 5% at 8 steps). Fine detail can differ a little. They are here so people can try them at 3.6 MB instead of 1.2 GB; for the originals, go to their authors' pages.

Settings:

CFG 1 · euler / simple · the model's shift 5 · strength 1.0, and

• 4steps v0.6 / v0.6 Diverse: 4 steps. At 4 steps they give 6–10% less colour than SFT at 25 steps at 1024 (6–14% at 2048); at 8 steps they are back near SFT's colour (only colour was measured at 8 for these two).

• 8steps v0.7: 8 steps, up to 1024 px.

• akhaliq cut: 4 steps (its author recommends 6–8 above 1024). lodestones cut: 8 steps.

Easiest in ComfyUI:

install HaDeS MOSS Suite from the ComfyUI Manager, plug a MOSS Turbo LoRA node into the turbo input of MOSS KSampler, set its steps (4 or 8), and it applies the LoRA with those settings. Stock nodes work too:LoraLoaderModelOnly + KSampler with the settings above.

Measured, not guessed.

Every number is in the Hugging Face card and in the suite's RESULTS.md, including where our LoRAs fall short (small text at 2048 x 2048; 8 steps at 2048).

• Our LoRAs on Hugging Face (files, model card, measurements): https://huggingface.co/DeVilDoNia/HaDeS-MOSS-Turbo

• HaDeS MOSS Suite (nodes, training script, every test): https://github.com/Devildonia/HaDeS-MOSS-Suite

Credits and licences:

• Base model: MOSS-Image SFT by the OpenMOSS Team (Apache 2.0).

• akhaliq — MOSS-Image-SFT-4-steps (Apache 2.0): https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps . Our DMD training recipe follows his; none of his weights are in our LoRAs. The cut here is his v2.0 file, reduced by us.

• lodestones — Kromoss and its 8-step DMD2 LoRA (Apache 2.0): https://huggingface.co/lodestones/Kromoss . His LoRA made for

• MOSS-Image SFT is at https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd (not the file cut here; not compared yet).

• Changes made to the two cuts: layers 24–41 only, each weight reduced to rank 2 by an exact SVD, saved in bf16. The source file, its SHA256 and these changes are written in each file's metadata.

Not tested:

other base models for our LoRAs (Base, finetunes, merges), other strengths, img2img quality.

Per-version texts ("About this version")

4steps - v0.6 — HaDeS-MOSS-Turbo-v0.6.safetensors · Base model: MOSS Image (SFT)

Ours. 4 steps, CFG 1, euler / simple, strength 1.0. Most robust on text at 4 steps.

4steps - v0.6 Diverse — HaDeS-MOSS-Turbo-Diverse-v0.6.safetensors · Base model: MOSS Image (SFT)

Ours. 4 steps, CFG 1, euler / simple, strength 1.0. Keeps more of SFT's variety on open prompts.

8steps - v0.7 — HaDeS-MOSS-Turbo-8steps-v0.7.safetensors · Base model: MOSS Image (SFT)

Ours. 8 steps, CFG 1, euler / simple, strength 1.0, up to 1024 px. Blind pairs at 8 steps: beat akhaliq's 4-step LoRA 18 to 5;

against lodestones' Base / Kromoss 8-step LoRA on SFT, 10 to 7 at 1024 and 0 to 5 at 2048.

akhaliq 4-step v2.0 (3.6 MB cut) — akhaliq-MOSS-SFT-4step-v2.0_L24-41_r2_cut.safetensors · Base model: MOSS Image (SFT)

Not ours: akhaliq's MOSS-Image-SFT-4-steps v2.0 (Apache 2.0), https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps , reduced by us to layers 24–41 at rank 2 (bf16). 4 steps, CFG 1, euler / simple, strength 1.0. His card recommends the unmerged fp32 original for the finest text.

lodestones Kromoss 8-step DMD2 (3.6 MB cut) — lodestones-Kromoss-8steps-dmd2_L24-41_r2_cut.safetensors · Base model: MOSS Image (Base)

Not ours: lodestones' 8-step DMD2 LoRA for MOSS-Image Base / Kromoss (Apache 2.0), https://huggingface.co/lodestones/Kromoss ,

reduced by us to layers 24–41 at rank 2 (bf16). 8 steps, CFG 1, euler / simple, strength 1.0. Made for Base / Kromoss; it also runs on SFT, where we tested it.

HaDeS