DeVilDoNia/HaDeS-MOSS-Turbo

HaDeS MOSS Turbo

Three few-step LoRAs for MOSS-Image SFT: two for 4 steps and one for 8 steps, all at CFG 1 instead of 25 steps at CFG 2.5. Each file is 3.4 MB (rank 2, layers 24–41 only), trained from zero on one RTX 5090.

File	What it is for
HaDeS-MOSS-Turbo-v0.6.safetensors	The most robust on text: layout JSON 6 / 6, a 16:9 three-line poster 3 / 3
HaDeS-MOSS-Turbo-Diverse-v0.6.safetensors	Keeps more of SFT's own spread on open prompts (faces, poses, props)
HaDeS-MOSS-Turbo-8steps-v0.7.safetensors	For 8 steps: in blind pairs it beat akhaliq's 4-step LoRA 18 to 5 and matched lodestones' Base / Kromoss 8-step LoRA (run on SFT) up to 1024

Use

• Model: MOSS-Image SFT (moss_image_sft_bf16 from Comfy-Org/MOSS-Image), text encoder and VAE as usual.
• Settings: CFG 1, euler / simple, the model's shift 5.0, LoRA strength 1.0, and 4 steps for Turbo and Turbo Diverse, 8 steps for Turbo 8-steps. These are the settings measured.
• At 1024 x 1024 and above, Turbo and Turbo Diverse prefer more steps for colour: at 4 steps they give 6–10% less colour than SFT at 25 steps (Lab chroma, 20 prompts at 1024; 6–14% on 5 prompts at 2048); at 8 steps Turbo is at 1.09 of SFT's colour (1.03 at 2048) and Turbo Diverse at 1.02 (0.91 at 2048). Only colour was measured at 8 steps for these two; their text and layout results below are at 4.
• ComfyUI, stock nodes: LoraLoaderModelOnly on the model, then KSampler with the settings above.
• ComfyUI, HaDeS MOSS Suite: a MOSS Turbo LoRA node into the turbo input of MOSS KSampler; it applies the LoRA and these settings for you (set its steps to 8 for Turbo 8-steps).

Measured

At 1024 x 1024 through the stock LoraLoaderModelOnly, euler / simple / 4 steps / CFG 1, against akhaliq's 4-step v2.0 LoRA (rank 128, all layers, 1.2 GB) at the same settings. Counts are by eye; ethnicity counts approximate. Full tables, scripts and every version tried: tools/RESULTS.md in the HaDeS MOSS Suite.

Test	akhaliq 4-step v2.0	Turbo v0.6	Turbo Diverse v0.6
Sign and three-line poster, 6 seeds each	12 / 12	12 / 12	12 / 12
Layout JSON, 6 seeds: strings right	5 / 6	6 / 6	5 / 6
Layout JSON: two-line title as its \n asks	0 / 6	4 / 6	1 / 6
1344 x 768 three-line poster, 3 seeds	2 / 3	3 / 3	0 / 3
1344 x 768 sign, 3 seeds	2 / 3	2 / 3	3 / 3
"photograph of a woman sitting in a café", 16 seeds: East Asian	~15 / 16	~14–15 / 16	~9–10 / 16
Explicit ethnicity prompts (3 x 6 seeds)	18 / 18	18 / 18	18 / 18
2048 x 2048 street, fisherman, sign; 768 x 1344 fisherman, sign, fox (3 seeds each)	clean	clean	clean
2048 x 2048 sign: small line "open 7am - 3pm" right	2 / 3	1 / 3	2 / 3
Seconds per image at 2048 x 2048 (RTX 5090, SageAttention, median of 9)	7.2	6.9	7.2

• On an open prompt MOSS-Image SFT at 25 steps gives about 10 East Asian women in 16 seeds; Turbo Diverse keeps that mix, Turbo and akhaliq's LoRA give almost only East Asian women. Name the ethnicity in the prompt for anything specific: all three follow it.
• At 2048 x 2048 the large lettering stays right on every file; small text is where they slip, Turbo more than the others here.
• The stock loader adds the LoRA into the weights before sampling, so no file adds work per step; the timing row is one run each, not repeated.
• Not measured: other base models (Base, finetunes, merges), strengths other than 1.0, step counts other than those above, img2img / denoise below 1.0, editing. They run (checked: an int8 finetune as the base, img2img at denoise 0.6), but their quality is untested.

Turbo 8-steps v0.7, at 8 steps

20 prompts in natural language (portraits, full-body poses, hands, animals, neon and chalkboard text, watercolor, anime, food, interiors, a product shot, a market, a band, landscapes), one seed each, at 1024 x 1024 and the first five at 2048 x 2048; stock LoraLoaderModelOnly + KSampler, 8 steps, CFG 1, euler / simple. Judged blind in pairs (left / right at random, a tie allowed):

Against, at 8 steps (full file)	1024: ours / theirs / tie	2048: ours / theirs / tie
akhaliq 4-step v2.0 (rank 128, 1.2 GB)	14 / 4 / 2	4 / 1 / 0
lodestones 8-step DMD2 for MOSS-Image Base / Kromoss, run on SFT (rank 128, 1.2 GB)	10 / 7 / 3	0 / 5 / 0

• Colour as a ratio to SFT at 25 steps (Lab chroma): Turbo 8-steps 1.01 at 1024 and 0.90 at 2048; lodestones' 1.11 and 0.99; akhaliq's 1.27 and 1.11 (both more saturated than SFT itself).
• At 2048 x 2048 that lodestones LoRA wins (5 of 5 pairs) and has more colour; Turbo 8-steps is for up to 1024.
• The lodestones file compared is 8-steps-dmd2-lora.safetensors from lodestones/Kromoss (SHA256 7683fe1f…), distilled from MOSS-Image Base; an earlier version of this card called it lodestones' 8-step LoRA without saying so. lodestones' SFT 8-step LoRA (lodestones/MOSS-Image-SFT-8-steps-dmd) has not been compared.
• At 4 steps Turbo 8-steps is not better than Turbo / Turbo Diverse (its café mix: ~13 East Asian women in 16 seeds).
• One seed per prompt, judged by eye. Not measured: the text tests above at 8 steps, other seeds, sizes between 1024 and 2048.

How they were trained

DMD (distribution matching distillation) with a regression term, tools/moss_dmd_train.py in the HaDeS MOSS Suite:

• one frozen bf16 copy of SFT; the generator LoRA (rank 2, layers 24–41, from zero) and a fake-score LoRA (rank 32, all layers) as fp32 hooks on it;
• teacher: SFT at CFG 3 with the negative pass skipping layer 22, as in akhaliq's recipe; generator trained on the euler / simple 4-step schedule with DMD2's backward simulation;
• regression term at weight 0.05, toward 256 images SFT rendered at 25 steps from fixed noises: it keeps the teacher's seed-to-image map, which DMD alone collapses to one composition per prompt;
• 512 x 512 only, 1,024 generated prompts, batch 1, lr 5e-5, 2,000 updates (about 45 minutes, plus 19 for the regression images). Turbo is update 1,600, Turbo Diverse update 1,000 of the same run.
• Turbo 8-steps v0.7: the same, plus DMD2's two time-scale rule (5 fake-score updates per generator update), 3,000 updates (3.4 s each); it is update 3,000.

Credits and license

Apache License 2.0 (see LICENSE and NOTICE). Trained from MOSS-Image SFT by the OpenMOSS Team (Apache 2.0). The training recipe follows akhaliq's MOSS-Image-SFT-4-steps (Apache 2.0); none of its weights were used.