MOSS-Image Derivative Checkpoints, LoRAs, and Workflows (November 2026)
Contents
When I previously evaluated MOSS-Image Base and SFT on an M1 Max Mac in October, only the two official releases existed.
Roughly two and a half weeks later, a wave of fine-tunes, merges, acceleration LoRAs, image-editing LoRAs, and ComfyUI workflows has landed across Civitai and Hugging Face.
In this context, checkpoints refer to standalone model weights, LoRAs are lightweight adapter files adding art styles or step reductions, and workflows are pipeline configurations connecting nodes in ComfyUI.
Following the approach in my October roundup of Anima derivatives, I surveyed what the community is distributing.
I have not downloaded or run generation tests on these models; the descriptions in the tables summarize what authors wrote on their respective release pages.
Survey Scope
On November 2, 2026, I collected Civitai listings through the Civitai API by specifying baseModels=MOSS Image, nsfw=true, sort=Most Downloaded, and period=AllTime, paging through to the final page for each category.
Data was fetched from civitai.red, and dates represent Civitai release dates in UTC.
| Type | Count |
|---|---|
| LORA | 1,602 |
| Checkpoint | 82 |
| Workflows | 68 |
| Other (prompt skills, custom nodes, etc.) | 18 |
| LoCon | 16 |
| DoRA | 2 |
| VAE | 2 |
| Total | 1,790 |
ControlNet, TextualInversion, and Upscaler had 0 entries.
Because this count covers only specified types, it does not represent every related file on Civitai; models where creators did not set the base model to MOSS Image are excluded.
Because a single model page often hosts multiple versions, there were 2,037 distinct versions tagged with MOSS Image. Of those, 2,032 appeared between October 17 and November 1—averaging roughly 127 versions per day.
Download counts represent the sum of stats.downloadCount for MOSS Image versions only.
Overall page totals include downloads for other base model architectures, so they were excluded from the tallies.
For instance, MiaoMiao RealSkin shows 66,037 downloads across its entire page, but its release for MOSS Image (named “MOSS 1.0”) accounts for 43. Similarly, Prime Lust has 7,381 total downloads across all architectures, with 1,923 attributed to its MOSS version. Given how recently the architecture debuted, download numbers strongly reflect differences in release timing.
Official Updates
The official weights have not changed since the October 16 release.
Checking commit histories on Hugging Face for Base and SFT shows only documentation revisions, link additions, and citation updates.
The primary updates in the official GitHub repository outside the core weights are summarized below:
| Date | Details |
|---|---|
| Oct 16 | Native MOSS-Image support merged into ComfyUI. SGLang and vLLM-Omni day-one support announced |
| Oct 18 | Community training toolkit KohakuMossTrainer featured in official README |
| Oct 19 | Technical report published on arXiv. Layout bounding box alignment fixes for text |
| Oct 20 | Support for fine-tuning and LoRA training announced in DiffSynth-Studio. Hosted service launched on MOSI API Platform |
| Oct 27 | Experimental Apple Silicon MPS support merged |
The MPS update relates to issues encountered during my previous M1 Max evaluation, where custom script modifications were needed to prevent float64 execution halts. Official macOS installation guidelines are now available, noting verified operation on an M1 Max 32GB system using SFT. Other Apple Silicon chips, configurations with lower unified memory, and the Base model are listed as unverified.
The official README continues to list only Base and SFT, with no image-editing variant.
According to thearkorbit’s model card, while official comments reference derivative work for image reference conditioning, the corresponding weights have not been published. Current image-editing LoRAs originate entirely from the community.
ComfyUI Integration
MOSS-Image runs natively in ComfyUI without custom nodes.
Initial support debuted in v0.42.0 (October 18), while v0.41.0 lacked support.
Color and depth conditioning, image-editing LoRAs, and dedicated MOSS model merging nodes arrived later in v0.43.0 (October 27), and were not backported to v0.42.1.
| Feature | Pull Request | Merge Date (UTC) |
|---|---|---|
| Native integration | #17095 | Oct 16 |
| Color and depth conditioning | #17206 | Oct 19 |
| Image-editing LoRA support | #17296 | Oct 22 |
| MOSS dedicated merge nodes | #17349 | Oct 25 |
Repackaged weights are hosted under Comfy-Org/MOSS-Image.
This collection includes BF16 and INT8 ConvRot editions of Base and SFT, text encoder variants in BF16, INT8, and w4a8, standard VAE files, and community color/depth control and image-editing adapters.
INT8 quantization options reduce storage overhead: based on Civitai API metadata, the INT8 model measures roughly 5.04 GiB compared to 9.47 GiB for BF16, nearly halving file footprint.
Official guidance recommends standard integration, relegating the repository’s bundled plugin to an optional fallback when loading directly from Hugging Face directory layouts.
Because the two approaches use different loader nodes and timestep scaling factors, existing workflow loaders cannot simply be swapped in place without rebuilding the graph.
Checkpoints
Among the 82 available checkpoints, top models ranked by downloads for their MOSS Image releases include:
| Model | MOSS Version DLs | Latest Release (UTC) | Author Description |
|---|---|---|---|
| MOSS Image (Civitai Mirror) | 21,325 | Oct 16 | SFT logged 17,429 and Base 3,896. Notes official recommendation: SFT for generation, Base for fine-tuning and LoRA training |
| PornMaster Moss | 12,746 | Nov 1 | Merged LoRA checkpoint for NSFW generation. Claims character LoRAs apply more readily over the Base architecture |
| Moody Moss Mix | 9,821 | Oct 29 | V3 rebuilds underlying LoRAs. Default parameters specify 25 steps at CFG 3 |
| Realism By Stable Yogi Moss | 9,666 | Oct 31 | Photorealism tuning. Distributed in 5 formats from Q4 GGUF to INT8. Recommends 10 steps, CFG 1, er_sde, Beta |
| Dirty-Moss | 5,845 | Oct 22 | Experimental candid photography style. Merges slider LoRAs and realism weights |
| MiaoMiao Harem MOSS | 5,282 | Oct 28 | Anime illustration style ported to MOSS-Image-Base, targeting text and layout capabilities |
| Jib Mix Moss Realistic | 4,311 | Oct 28 | Photorealism tuning with integrated 8-step LoRA; V2 recommends 12 steps |
| Moss Image INT8/INT6/INT4 | 2,917 | Oct 17 | Quantized versions of official weights. INT8 measures ~5.4 GB, INT6 ~4.1 GB |
| Kromoss | 2,771 | Oct 19 | Base full fine-tune primarily addressing default blurriness on Base |
Fine-Tunes and Merges
Most leading checkpoints represent LoRA merges.
Only four releases clearly state that they underwent full fine-tuning on fresh training data:
| Model | MOSS Version DLs | Author Description |
|---|---|---|
| Kromoss | 2,771 and 1,723 (Two Civitai pages) | Trained from Base at 1024px. 11,000 update steps across ~704k images. Bundles 8-step DMD and DMD2 distillation LoRAs |
| MOSS Cat Tower | 1,752 | Refines default anime styling. Recommends 50 steps, CFG 3, and Skip Layer Guidance bypassing layers 22, 24, and 26 |
| Moss Image FinalCut | 1,715 | Trained on both descriptive captions and short WD14 tags. 12–25 steps, sgm_uniform, er_sde |
| MossAD Demo | 374 | Prototype trained on 3,685 images across 57 illustrators from Danbooru, Gelbooru, and E621 to test artist styles, tags, and layouts |
MOSS-Image 2.5D Anime (2,321 downloads) is an experimental merge of stylistic LoRAs, while CottonMoss (776 downloads) combines several adapters.
CottonMoss blends an 8-step LoRA at reduced weight, suggesting 12 steps during inference. Its documentation notes that prompt adhesion can soften and consistent depiction of copyrighted characters remains difficult.
Notable creators from my October Anima derivative survey have also released MOSS models: nuko_masshigura (Anima Cat Tower) published MOSS Cat Tower, uraoto (CottonAnima) released CottonMoss, and MIAOKA (MiaoMiao series) shared MiaoMiao Harem MOSS.
Acceleration LoRA Derivatives
Official default parameters recommend 50 steps at CFG 3. Because CFG evaluates conditioning twice per step (conditional and unconditional), inference requires calculating the Diffusion Transformer (DiT) backbone 100 times per image.
This computational load explained why generating a single image took roughly 20 minutes on an M1 Max in my earlier benchmarks. Distillation LoRAs designed to cut step counts appeared on October 16, the day of MOSS’s release.
The lineage described across model cards unfolds as follows:
flowchart TD
A["MOSS-Image-SFT<br/>50 steps, CFG 3"] --> B["lodestones 8 steps<br/>data-space DMD"]
B --> C["darask0 4 steps<br/>DMAD"]
C --> D["akhaliq 2 steps<br/>IDM"]| LoRA | Steps | Author Description |
|---|---|---|
| lodestones/MOSS-Image-SFT-8-steps-dmd | 8 | No CFG required. Marked work-in-progress. Recommends weight 1.0 for lora.safetensors and 0.7 for lora_latest. Base variant hosted in a separate repository |
| darask0/MOSS-Image_4-step_adapter_lora | 4 | DiT compute time for 1024x1024 on H100 dropped from 16.2s (50 steps) to 0.81s (excludes text encoder and VAE). Documents known visual regressions |
| akhaliq/MOSS-Image-SFT-4-steps | 4 | Published on Civitai as a v2 iteration compatible with standard Euler sampling |
| akhaliq/MOSS-Image-IDM-2-Steps | 2 | Recommends export 750 from v1 for text-to-image. The author notes v2 tuning for editing degraded pure generation quality |
| Coffmeteor/MOSS-Image-Turbo | 16 | Beta release operating at CFG 1. Slightly softens skin glossiness, but notes anatomy and multi-character composition remain uncorrected |
On Civitai, the top-downloaded LoRA overall was Moss Image Turbo LoRas (18,592 downloads), which bundles several of these community distillations.
Beyond step pruning, p1atdev/MOSS-Image-SFT-KV-Cache-LoRA focuses on caching and reusing text encoder projections. The author’s benchmarks on an RTX 6000 Ada showed that a long 2,941-token layout prompt dropped from 8.03s to 5.26s, though a short 53-token prompt marginally slowed from 3.98s to 4.39s.
Optimization and Bug Fixes
| Release | Target | Author Description |
|---|---|---|
| unsloth/MOSS-Image-SFT-GGUF | GGUF Quantization | Runs on CPU, Mac, and GPU via stable-diffusion.cpp. Recommends UD-Q4_K_XL (3.42 GB). Qwen3.5-9B text encoder must be supplied separately |
| KBlueLeaf/MOSS-Image-FP16-fixed | FP16 overflow | Scales internal activations by 1/16 within model weights, ensuring 16x headroom in FP16 to mirror FP32 stability |
| MATLOWAI/MOSS-Image-SFT-DeBand | SFT vertical banding | LoRA suppressing faint vertical stripes appearing in uniform sky and wall regions by reverting select SFT layers to Base weights |
| nynxz/MOSS-Image | SFT vertical banding | 1.9 MB SFT Destripe LoRA. DeBand documentation references it as the least visually disruptive artifact patch |
| comfyanonymous/MOSS_image_MAX_ULTRA_hires_lora_test | 4MP+ rendering | Experimental LoHa adapter enabling high-resolution generation above 4 megapixels on Base |
| PotatoForge/moss-image-4b-encoder-adapter | Text encoder memory | ~21 MB projection adapter mapping Qwen3.5-4B outputs into 9B dimension spaces. Disclaims full parity with native 9B outputs |
| mlx-community/MOSS-Image-SFT-bf16 | Apple Silicon MLX | MLX conversion for native macOS framework execution |
The community showcase in the official README lists DiffSynth-Studio, DeBand, lodestones’ 8-step LoRA, FP16-fixed, and ComfyUI-Spectrum-MOSS.
A censorship-bypass text encoder, Heretic, has also appeared.
According to its author, when standard Qwen3.5-9B models handle prompt expansion, adult prompt tokens often trigger refusals that get embedded directly into generation prompts, disrupting downstream image rendering. Heretic seeks to prevent these refusals.
Image Editing and Reference Conditioning
As observed during my October testing, official MOSS-Image pipelines handle only text-to-image without native image input.
However, thearkorbit’s documentation highlights that one of MOSS’s four spatial coordinate slots sits unused during standard text generation. Injecting reference image latents into this unused slot allows the model to recognize reference visual features without re-architecting the base network.
Because this injection primarily copies visual features rather than interpreting transformation prompts, companion LoRAs are trained to teach instruction following. This technique was pioneered in ostris’s ai-toolkit.
| Release | Purpose | Author Description |
|---|---|---|
| chinoll/Moss-Image-Edit | Prompted editing | Alters color, clothing, lighting, seasons, and artistic styles across Base and SFT, supporting up to 4 reference images. Comfy-Org hosts a rank 64 build |
| akhaliq/MOSS-Image-Edit-HQ | Prompted editing | Higher-fidelity branch targeting improvements over chinoll. Recommends armV5; notes armV6 did not show quality gains |
| 7err4/MOSS-Image-CharRef | Character reference | Renders a character from a single reference image into new scenes. Base-only and anime-focused; uses 50 steps at CFG 3 |
| pleiades7/MOSS-Image-CharRef-PoC | Character reference | Precursor prototype to CharRef. Openly notes facial geometry and hairstyle consistency frequently drift. Also mirrored on Civitai |
| Anime2Real | Anime-to-photorealism | Image-to-image editing LoRA trained via ai-toolkit, with an accompanying workflow. Author states it does not yet match Krea2 quality |
| darask0/MOSS-Image-color-depth-control | Layout and pose control | Conditions generation on color-blocked segmentation and depth maps, targeting pose placement from 3D mannequins. Natively integrated into ComfyUI |
| Kaloscope-style-adapter | Style conditioning | Blends artist style signatures or applies reference styles. Compatible with Krea2 and MOSS-Image-Base |
| paralaif/MOSS-Image-Alpha-LoRA | Transparent PNG | Experimental adapter outputting alpha channels paired with specialized RGBA VAEs. Trained on 39 images |
CharRef and its PoC prototype require ostris’s custom nodes.
The author of Anime2Real advises using updated node versions after recent ComfyUI backend changes, meaning node versions must be matched carefully against specific ComfyUI releases.
Workflows
Looking at the 68 workflows ranked by downloads for their MOSS Image versions, top entries focus on minimal pipelines combined with step acceleration.
| Workflow | MOSS Version DLs | Details |
|---|---|---|
| Moody Moss Minimal | 5,561 | Provides dual setups for standard 25-step generation and accelerated inference |
| PornMaster Moss SFT t2i | 4,634 | Text-to-image workflow tailored for PornMaster |
| Text to Image, 4-Step Turbo + SeedVR2 2x | 4,068 | Combines 4-step acceleration with SeedVR2 2x upscaling |
| Simple Workflow (T2I BBOX, I2I) | 3,786 | Bounding-box layout placement, depth-assisted image-to-image, and second-pass refining |
| MOSS Image inpainting workflow | 1,851 | Inpainting setup |
| MOSS Image workflows | 1,615 | Multi-feature setup covering ADetailer face/hand repair, upscaling, inpainting, prompt expansion, editing, and depth |
| Anime2Real Workflow | 1,555 | Pipeline for converting anime illustrations into photorealistic renderings |
| MOSS 4-step workflow | 1,424 | 4-step distilled generation setup |
| MOSS Pro Grade | 1,273 | Modular setup adding editing, control maps, inpainting, and weighted prompting across versions |
| T2I + LF Block Cache | 1,251 | Text-to-image pipeline integrating LF Block Cache |
| Artfat MOSS BBox Studio | 1,204 | Canvas bounding boxes, editing nodes, and LLM-assisted prompt generation |
Sorting by overall page downloads elevates legacy collections such as a multi-architecture workflow pack (38,863 total) and an inpainting suite originating from SD 1.5 (23,696 total), but filtering strictly to MOSS-specific versions leaves only 18 and 605 downloads respectively.
MOSS’s standout capability remains its JSON layout parsing. Interactive tools like BBOX Prompter, which exports layout coordinates from canvas drawings, and AI Prompt Canvas, originally built for Ideogram 4, now offer dedicated MOSS layout modes.
Notable specialized setups include:
| Workflow | Purpose | Author Description |
|---|---|---|
| Regional character LoRAs | Dual-character LoRA separation | Because native layout boxes modulate only prompts, this workflow generates a base image on pass 1, then inpaints isolated character LoRAs inside targeted bounding boxes on pass 2 |
| MOSS Doujin Kit | Manga book generation | Experimental pipeline generating complete multi-page manga; companion standalone tool Fumika also exists |
| MOSS LoRA Trainer | Single-node LoRA training | Prototype running LoRA fine-tuning inside a single ComfyUI node |
| Mac向けSeedhunt | Apple Silicon inference | Tested on M5 Pro 48GB and M1 MacBook Air 16GB. Recommends 25 steps, CFG 2.5, er_sde, and simple scheduling |
| 8GB VRAM 2.5D Anime | Low VRAM execution | Generates 1024x1536 in ~62 seconds on an 8GB laptop GPU without external custom nodes |
For prompting assistance, besides the official Prompt Enhancer, authors distribute model-tailored LLM system prompts such as Moody Moss Prompt Writing Skill.
LoRA Distribution
LoRA, LoCon, and DoRA listings combined for 1,620 models, accumulating 453,821 downloads across MOSS Image versions.
Among the top 50 models, 23 carry Civitai’s NSFW content flag.
Leading categories focus heavily on photorealism and anatomy sliders. Prominent photorealism entries include Lenovo UltraReal (8,213 downloads), NiceGirls UltraReal (6,663 downloads), and Phone Photography (5,033 downloads).
On the stylized side, Anireal (2,442 downloads), Anime Flat Style (2,075 downloads), and Hentai Style (1,463 downloads) rank near the top.
Many creators adapted established LoRA lines by publishing new Moss version files alongside existing SDXL or Illustrious checkpoints.