AiO Generator (MOSS)

MOSS-Image is new enough that the local scene hasn't caught up. I went looking for community threads and found essentially nothing - the model is barely named outside the pack that supports it. If you're running MOSS, you're near the front of the queue, and you should expect to be the person who finds the rough edges.

Which is why a node that runs the whole chain is a decent place to start. AiO Generator (MOSS) does encoding, sampling, highres, detailing, upscaling and saving off one wire - the MOSS member of the three-generator family in ComfyUI-EasyUseAnima by n0va39.

What it is

An output node that takes one context input and gives you a finished image plus its metadata. Underneath, nothing exotic: it loads your MOSS diffusion model, Qwen3.5 9B text encoder and FLUX.2 VAE from the settings in Easy Use Anima Input, builds an empty latent, calls core KSampler, optionally re-samples for Highres, optionally runs a SAM3 face/eye detailer, optionally upscales with USDU or ResShift, then saves with prompt and workflow embedded. The value is setting it up once instead of forty times.

It's fixed-family: this is the MOSS node, it won't run Krea 2 or Anima, and settings live in a versioned JSON string the popup UI edits for you. Don't hand-edit that string, and don't expect an Anima profile to load.

On the pack: one maintainer, actively shipping (changelog 1.6.0 is where these separate generators landed), GPL-3.0-or-later, and a very thin public footprint. No threads, no tutorials, nobody to ask. Read the release notes; they're unusually detailed and the closest thing to documentation this corner has.

How it works

MOSS here is native MOSS: ComfyUI has to know what MOSS is. That means CLIP type = moss_image in the CLIP loader and the FLUX.2 latent node, which is why the pack demands ComfyUI 0.42.1 or newer. The generator builds its empty latent via EmptyFlux2LatentImage specifically - MOSS lives in FLUX.2's latent space, so dimensions must be multiples of 16.

The defaults tell you this isn't an 8-step distilled model: 25 steps, CFG 2.5, er_sde/simple, shift 5.0, applied through the AuraFlow model-sampling patch. er_sde is an odd-looking default for the family and a very normal one for Qwen-encoder DiTs, so the author picked the sampler that behaves rather than the obvious euler.

The feature set is deliberately narrow. Available: Highres, Detailer, USDU/ResShift upscale, Artist Mix, SageAttention, FP16 accumulation, generic AiO hooks, and NegPip if you have a native-MOSS-capable ComfyUI-ppm. Unavailable: Spectrum forecasting (unless a separate provider exposes the exact NativeMossSpectrumPatch contract), Mod Guidance, DAVE, Safe PAG, DiT corrections, and Torch Compile - the last because, per the release notes, the current native attention and quantized kernels fail to compile. All of it is locked in the UI rather than silently ignored.

Inputs and outputs that matter

Required: easy_use_anima_input. Point it at your MOSS-Image diffusion model, a Qwen3.5 9B encoder, CLIP type = moss_image and a FLUX.2 VAE. That's the whole setup, and getting one of those four wrong is the most common way this node fails to start.

Required but ignore it: generation_settings, the hidden blob.

Optional: lora_stack - applied to MODEL and CLIP before conditioning, so it behaves like any LoRA stack - and aio_hook, empty unless a third-party extension supplies one.

Three outputs. image is the decoded result; the generator already saves WebP into EasyUseAnima/AiO by default, so adding your own SaveImage gets you duplicates. latent lets you decode with another VAE or keep sampling. metadata_json is the summary of what actually ran - the thing you want when a generation is wrong and you can't tell whether Highres fired.

Install

Manager: search ComfyUI Easy Use Anima. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/n0va39/ComfyUI-EasyUseAnima

Restart ComfyUI, hard-refresh the browser. Python deps are light (requests plus googletrans-py), so the install is rarely where this breaks. The weights are the problem: MOSS diffusion model, Qwen3.5 9B encoder, FLUX.2 VAE. Nothing is downloaded or substituted for you, so a half-populated models folder gives you a node that looks fine and errors at queue time.

Optional packs, only for the extra stages: ComfyUI-KJNodes for SageAttention and FP16 accumulation, ComfyUI-Impact-Pack for the SAM3 detailer, ComfyUI-Spectrum-KSampler if you're chasing forecasting. Without them the relevant toggles lock instead of crashing your graph.

Where people get burned

The loudest error is MOSS requires ComfyUI with EmptyFlux2LatentImage (tested v0.42.1) - that's a stale host, and reinstalling this pack won't fix it. Update ComfyUI.

MOSS requires CLIP type 'moss_image' means the CLIP loader is on the wrong type; AiO Generator (MOSS) requires ... means the checkpoint in Input isn't a MOSS model. Both are the family check catching a mis-wire, not a bug.

Then the quiet one: don't import distilled-model habits. Carry 8 steps and CFG 1 over from a Turbo model and MOSS hands you mush, and you'll blame the node. Start at 25 and 2.5 and move one number at a time.