Eric UniPic3 Load Model

Before anything else: check your VRAM. Eric UniPic3 Load Model wants roughly 58GB in bf16 - a ~41GB transformer, a ~16.5GB Qwen2.5-VL text encoder, and a VAE - and the README's recommendation is a 64GB+ card. If your GPU is consumer-sized, this is the node where the pack stops being an idea and becomes an out-of-memory error. If you've got the silicon, it's also the cleanest way to get the whole model in and out of memory.

What it does

Every workflow in this pack starts here. The loader builds the complete UniPic3 pipeline - transformer, VAE, text encoder, scheduler - into one object and hands it to the Edit and Compose nodes over a single wire. Nothing else runs until this node succeeds, because it's doing all the heavy lifting: finding the weights, placing them on your device, and keeping them warm for the next run.

How it works

Under the hood it's the diffusers QwenImageEditPlusPipeline - the same machinery Qwen-Image-Edit-2511 uses - with Skywork's own ~13B QwenImageTransformer2DModel (60 layers, 24 attention heads) swapped in for Qwen's. The loader pulls the transformer from a local folder, builds the full pipeline around it from your base pipeline path, moves it to your chosen device, and enables VAE tiling so high-res decoding doesn't blow up memory. Two things are worth knowing before you blame it for failing:

• Everything loads with local_files_only=True. This node will never download a model for you. If the folder you pointed at doesn't contain the weights, it raises a ValueError, not a fetch.
• The pipeline is cached globally, keyed on variant plus transformer path. Loading the same variant twice is near-instant, and switching variants automatically evicts the old one first.

There are three variants because Skywork distilled the model three ways: teacher (the full 50-step model, ~30-60s per image), dmd (8 steps, ~5-10s, close quality), and consistency (≤8 steps, geometry-focused, weaker on some edits). The 12.5x speedup is the usual distillation bargain - fewer steps means less time to refine detail, so the fast variants are for iteration, not finals.

The inputs that matter

• variant - teacher / dmd / consistency.
• base_pipeline_path - where your Qwen-Image-Edit-2511 download lives.
• transformer_path_override - where the Skywork transformer weights live; leave empty to use the variant default.
• precision - bf16 is the default and the right call on RTX 40/50 series; fp16 if your card can't do bf16.
• keep_in_vram - True caches the pipeline for fast re-runs; False frees it every time at the cost of a full reload.

The output is a single pipeline wire, typed UNIPIC3_PIPELINE, which feeds into Eric UniPic3 Image Edit or Compose (HOI).

Installing it

ComfyUI Manager (search "Eric UniPic3") or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Eric_UniPic3
pip install -r Eric_UniPic3/requirements.txt

That only gets you the code. The README then expects four repos - Skywork/Unipic3 (teacher), Skywork/Unipic3-DMD, Skywork/Unipic3-Consistency-Model, and the Qwen/Qwen-Image-Edit-2511 base - about 100GB of disk. Grab them with huggingface-cli download <repo> --local-dir <your/path>, restart ComfyUI, and you'll see all four nodes under "Eric UniPic3". Dependencies are light (diffusers ≥0.35, transformers ≥4.45, Pillow); the heavy lifting is the weights.

Common issues

• "Transformer path not found for variant 'teacher'" - the default paths are the author's own Windows H:\ drive, and on any other machine those folders don't exist. Set base_pipeline_path and transformer_path_override to your real locations. This is the #1 error, and it's not your fault; it's a hardcoded dev-machine path.
• Out of memory - 58GB is 58GB. If you're near the edge, dmd + keep_in_vram off + bf16 buys a little headroom, but you're in server-GPU territory regardless.
• Slow first load - you're moving ~58GB of weights. The cache makes run two instant.

The honest take

You're early to this party. UniPic3 shipped in January 2026 and there's essentially no community chatter about it yet - search the usual places and you find UniPic 1 and 2, not 3. Fewer tutorials, sure, but also fewer people yelling at you. And the README's own future-work list concedes that the real accessibility fix is quantization (INT8, NF4, GGUF) that doesn't exist yet. For now this pack is for big-GPU owners, and this loader is the honest, well-documented way in.