<h1 id="few-step-moss-image-in-3.4-mb.">Few-step MOSS-Image in 3.4 MB.</h1>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.<h3 id="ours-(hades):">Ours (HaDeS):</h3><ul><li>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).</li><li>4steps - v0.6 Diverse: keeps more of SFT's own variety on open prompts (faces, poses, props).</li><li>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).</li></ul><h3 id="not-ours-(credited-cuts-apache-2.0):">Not ours (credited cuts, Apache 2.0):</h3><ul><li>akhaliq 4-step v2.0 (cut): akhaliq's MOSS-Image-SFT-4-steps v2.0, reduced to layers 24–41 at rank 2.</li><li>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.</li></ul>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.<h3 id="settings:">Settings:</h3>CFG 1 · euler / simple · the model's shift 5 · strength 1.0, and<ul><li>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).</li><li>8steps v0.7: 8 steps, up to 1024 px.</li><li>akhaliq cut: 4 steps (its author recommends 6–8 above 1024). lodestones cut: 8 steps.</li></ul><h3 id="easiest-in-comfyui:">Easiest in ComfyUI:</h3>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:<code>LoraLoaderModelOnly</code> + <code>KSampler</code> with the settings above.<h3 id="measured-not-guessed.">Measured, not guessed.</h3>Every number is in the Hugging Face card and in the suite's <code>RESULTS.md</code>, including where our LoRAs fall short (small text at 2048 x 2048; 8 steps at 2048).<ul><li>Our LoRAs on Hugging Face (files, model card, measurements): <a target="_blank" rel="ugc" href="https://huggingface.co/DeVilDoNia/HaDeS-MOSS-Turbo">https://huggingface.co/DeVilDoNia/HaDeS-MOSS-Turbo</a></li><li>HaDeS MOSS Suite (nodes, training script, every test): <a target="_blank" rel="ugc" href="https://github.com/Devildonia/HaDeS-MOSS-Suite">https://github.com/Devildonia/HaDeS-MOSS-Suite</a></li></ul><h3 id="credits-and-licences:">Credits and licences:</h3><ul><li>Base model: MOSS-Image SFT by the OpenMOSS Team (Apache 2.0).</li><li>akhaliq — MOSS-Image-SFT-4-steps (Apache 2.0): <a target="_blank" rel="ugc" href="https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps">https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps</a> . 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.</li><li>lodestones — Kromoss and its 8-step DMD2 LoRA (Apache 2.0): <a target="_blank" rel="ugc" href="https://huggingface.co/lodestones/Kromoss">https://huggingface.co/lodestones/Kromoss</a> . His LoRA made for</li><li>MOSS-Image SFT is at <a target="_blank" rel="ugc" href="https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd">https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd</a> (not the file cut here; not compared yet).</li><li>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.</li></ul><h3 id="not-tested:">Not tested:</h3>other base models for our LoRAs (Base, finetunes, merges), other strengths, img2img quality.Per-version texts ("About this version")4steps - v0.6 — <code>HaDeS-MOSS-Turbo-v0.6.safetensors</code> · 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 — <code>HaDeS-MOSS-Turbo-Diverse-v0.6.safetensors</code> · 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 — <code>HaDeS-MOSS-Turbo-8steps-v0.7.safetensors</code> · 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) — <code>akhaliq-MOSS-SFT-4step-v2.0_L24-41_r2_cut.safetensors</code> · Base model: MOSS Image (SFT)Not ours: akhaliq's MOSS-Image-SFT-4-steps v2.0 (Apache 2.0), <a target="_blank" rel="ugc" href="https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps">https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps</a> , 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) — <code>lodestones-Kromoss-8steps-dmd2_L24-41_r2_cut.safetensors</code> · Base model: MOSS Image (Base)Not ours: lodestones' 8-step DMD2 LoRA for MOSS-Image Base / Kromoss (Apache 2.0), <a target="_blank" rel="ugc" href="https://huggingface.co/lodestones/Kromoss">https://huggingface.co/lodestones/Kromoss</a> ,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

HaDeS MOSS Turbo

akhaliq-MOSS-SFT-4step-v2.0_L24-41_r2_cut.safetensors

Files with SHA256: 2be3aaee3b42f886f8942ed78a0fde95b1d7cc46d47143f822f6b0866b3e95f0