Eric Qwen-Edit Load Model

Every node in the edit half of this pack starts here. Eric Qwen-Edit Loader loads the real Hugging Face QwenImageEditPlusPipeline from a local directory and hands it out as a single QWEN_EDIT_PIPELINE wire that every edit node - Image, Inpaint, Multi-Image, Style Transfer, Unify, Spectrum - expects on its left side. No pipeline, no editing.

The important thing to understand before you run it: this is not a ComfyUI-native loader. It doesn't produce a MODEL, a CLIP, or a VAE that you'd feed into a KSampler. It loads the whole diffusers pipeline in one shot, which is the entire point of this pack. Qwen-Image-Edit is a 20B MMDiT that the stock diffusers pipeline force-crushes to ~1 MP outputs no matter what you feed it; Eric's loader patches the pipeline to preserve your input resolution (aligned to 32 px) up to a configurable cap. That's why people reach for this pack instead of the generic "load diffusion model" route - the default ComfyUI loaders produce wrappers these nodes won't accept, and you lose the resolution preservation.

What you actually set

• model_path - path to a full Qwen-Image-Edit model directory (Qwen/Qwen-Image-Edit-2511 or 2509 from HuggingFace). It needs the folder - model_index.json, scheduler config, tokenizer, processor, and the ~54 GB of sharded safetensors - because the diffusers from_pretrained path reads all of it. The default value in the widget is just Eric's own Windows training path; you must replace it. There's no GGUF shortcut here; this loader is all-bf16.
• precision - bf16 is the default and the right call on any RTX 40/50 card. fp16 if you're pinched, fp32 if you enjoy watching your GPU smoke.
• keep_in_vram - True by default, and you want it on if you're iterating; the loader caches the pipeline so your second run doesn't re-read 54 GB from disk.
• offload_vae, attention_slicing, sequential_offload - the VRAM triage. Offload the VAE to CPU (saves ~1-2 GB), slice attention (smaller peak, slower), or go full sequential CPU offload (very slow, survives huge images). On a 24 GB card you shouldn't need any of them at 2 MP; on an 8 GB card, sequential_offload is the difference between running and OOM - one user on a 4060 found it mandatory, and that disabling it OOMs even with the other toggles on.
• vae_tile_blend - cosine (default) or linear, controlling how tiled VAE decode seams are blended. Cosine kills faint grid lines on gradients; leave it alone unless you see seams.

The output is one pipeline socket (QWEN_EDIT_PIPELINE). Feed it straight into the edit node of your choice; you can also chain Eric Qwen-Edit Spectrum in between for a speedup, and Eric Qwen-Edit LoRA before that to load adapters.

Install and the honest VRAM talk

Install is the same as every node in this pack:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments.git

or search "Eric Qwen Edit" in ComfyUI Manager and hit Install. Then download the model to disk - that's the real install step, and it's ~54 GB for the full bf16 edit model. The README's VRAM table is not marketing: 24 GB handles up to ~2 MP, 48 GB up to ~6 MP, 96 GB up to ~16 MP. Eric himself runs an RTX 6000 Pro, so "it's slow" reports from people on 5090s are not a surprise.

Two gotchas worth knowing before they bite. First, the pack pins diffusers<0.38 and transformers<5 in its requirements - if another pack forces a newer diffusers, this loader can break with a HybridCache import error, and vice versa. Second, this pack is dual-licensed CC BY-NC 4.0 (commercial license available on request) - fine for personal and hobby use, a licensing conversation you need to have before you sell anything made with it. Start with max_mp low (the edit node's cap), confirm the loader holds, then push resolution.