ND911/Verdigris_MOSS

Verdigris MOSS

A merged, requantized build in the MOSS-Image family, tuned toward character and fantasy work and fast enough for 8-step sampling workflows.

What this is

Verdigris is a merge, not a from-scratch fine-tune. It was built by folding earlier builds of this same lineage back into themselves, together with community LoRAs, over a long series of iterations. It inherits the MOSS-Image architecture (42-layer DiT, Qwen3.5-9B text encoder, FLUX.2 VAE) and is a drop-in replacement for a MOSS-Image checkpoint in ComfyUI.

About the lineage: it is not fully reconstructable, and this card does not pretend otherwise. The chain runs back through many intermediate merges, several distilled (DMD) variants and a stack of LoRAs applied at different stages over months, and some of those intermediate files no longer exist. What can be said with confidence:

• the strongest single ancestor is Kromoss by lodestones, itself a full fine-tune of MOSS-Image-Base by OpenMOSS;
• the rest is accumulated merge material and community LoRAs, mixed at varying weights;
• no attempt is made to claim reproducibility of the exact weights.

Treat it as a curated artifact, not a recipe.

Files

file	size	notes
diffusion_models/verdigris_moss_bf16.safetensors	10.2 GB	bf16, ComfyUI single file, fused proj.qkv
diffusion_models/verdigris_moss_int8_convrot.safetensors	5.4 GB	int8 W8A8 ConvRot, 256-wide groups, Comfy-Org layout
workflows/verdigris_moss_wf.json	216 KB	the reference workflow (ComfyUI)
examples/	—	nine example outputs

The int8 build quantizes the 168 block linears (proj.{qkv,o,ffn_up_gate,ffn_down}) with a regular Hadamard rotation per 256-wide input group and per-row scales; everything else stays bf16. Measured on this build: worst per-layer cosine 0.999952, i.e. about 1% relative weight error. The int8 file also carries per-block comfy_attention.config entries, so ComfyUI uses its int8 attention path for those blocks.

Requirements

• ComfyUI with MOSS-Image support (native, master/nightly).
• Text encoder: Qwen3.5-9B in models/text_encoders/, CLIP type moss_image.
• VAE: the full FLUX.2 VAE (flux2_moss_vae.safetensors in the reference workflow) in models/vae/.
• The workflow also references fibo2_dev_int8_convrot.safetensors (with a Qwen3-VL-4B text encoder and its own VAE, flux2_full_encoder_small_decoder.safetensors), a MoGe-3 depth model, and the DLSS5 and RTX VSR node packs. Those are only needed for the optional stages described below.

Usage

1. Drop the bf16 or int8 file into models/diffusion_models/.
2. Open workflows/verdigris_moss_wf.json. Its loaders already point at verdigris_moss_bf16.safetensors and verdigris_moss_int8_convrot.safetensors — pick whichever you downloaded.

The workflow is the reference for sampling settings: low step counts with manual sigma schedules, Skip Layer Guidance at layer 22 (24 in the refiner), a MOSS generation pass, an optional fibo2 touch-up pass at low denoise, a MOSS refinement pass, then RTX VSR and DLSS5. Do not treat the step counts or CFG values in isolation — they are tuned together with the sigma schedules and the guidance layers, and they will not transfer to a plain Euler/SGM sampler.

Which file to use

• bf16 — the reference. Use this first, and for anything that looks off.
• int8 — ~half the VRAM and disk, visually matched to bf16 on the frames checked so far.

Scope of testing — please read

This build has only been validated inside the workflow provided here — in ComfyUI, with the MOSS text encoder and full FLUX.2 VAE that the workflow uses.

Within that workflow, a range of sampler / scheduler / step-count / CFG combinations has been tried, and images both above and below the ~2 MP of the examples have been generated. Those runs are not documented here, so the shipped settings are the only ones with results you can compare against.

Not validated at all: diffusers or any non-ComfyUI runtime; the MOSS edit and control adapters; third-party LoRAs stacked on top; img2img, inpainting or upscale-refine chains. Results outside the workflow are possible but unmeasured.

Known weaknesses. Hands under occlusion and foreshortening (open palms, interlocked fingers) still fail; small man-made objects tend to have the right silhouette and no internal structure (a watch will have a band and no dial); legible text rendering is untested; very fine repeating detail (distant foliage, crowds) can tile.

Example images were generated at roughly 2 megapixels (1768×1131 and similar). Larger and smaller sizes have been generated but are not shown here.

fibo2-dev touch-up — the shipped workflow includes an experimental fibo2-dev pass (a second model used to re-render detail at low denoise) that is still being integrated and tuned. The examples in this repo should not be assumed to represent that stage at its final settings.

Examples

More in examples/.

Content note

The merge incorporates community adapters that include NSFW material, and the model is capable of producing NSFW output. It is unrated: treat it as 18+ and apply your own judgement and filters. Nothing here is filtered or safety-tuned.

License and credits

Free to use, including commercially. Released under Apache-2.0, following MOSS-Image and Kromoss — keep the attribution and the license notice and you are set. No gating, no permission needed, no restrictions beyond that.

• MOSS-Image — OpenMOSS, architecture, base weights, text encoder and VAE.
• Kromoss — lodestones, the fine-tune this build leans on most heavily.
• Numerous community LoRA authors whose adapters were merged in along the way.

This is an independent community build. It is not affiliated with, endorsed by, or supported by OpenMOSS or lodestones, and any defects are the merge's own.