Eric Qwen-Edit Image

This is the workhorse of the pack - the node you'll use for nine out of ten edits. Eric Qwen-Edit Image takes an image and an instruction ("change the background to a sunset", "make the person smile", "remove the text") and produces an edited image using the 20-billion-parameter Qwen-Image-Edit model. No mask, no inpainting dance, just a sentence.

The headline feature is resolution. The stock diffusers QwenImageEditPlusPipeline forces every output to ~1 MP - feed it a 12 MP photo and it comes back crushed. This pack patches the pipeline to preserve your input resolution (aligned to 32 px) up to a configurable max_mp cap. A 6 MP edit stays 6 MP, up to 16 MP if your VRAM allows. That alone is why people reach for this pack over the native ComfyUI Qwen nodes.

How it works

Where ComfyUI's native Qwen support decomposes the model into UNET → Scheduler → Sampler and forces you to hand-wire "Aura Flow Shift" values, this pack calls the real Hugging Face QwenImageEditPlusPipeline end-to-end with the FlowMatchEulerDiscreteScheduler. Every sigma shift, timestep, and conditioning step matches what the model was trained with, and the resolution-aware time-shift μ is interpolated from your output size automatically. Zero shift-node configuration.

Qwen-Edit also uses true CFG: two full transformer forward passes per step (conditional + unconditional). That's why the true_cfg_scale input is the main quality control and why it costs roughly double the compute per step.

Inputs that matter

• pipeline - from any pack loader.
• image - what you're editing. prompt - the instruction, written as a sentence.
• steps - 8 for a Lightning LoRA, 50 for the base model. This is the single biggest speed lever.
• true_cfg_scale - 1.0–20.0, default 4.0. Lower = less prompt adherence, more "creative drift".
• max_mp - output ceiling (0.5–16, default 8). Input is preserved up to this cap. Drop it to 4–6 while iterating, then raise for the final render - the README's tip #1.
• upscale_to_max_mp - when on, small inputs get upscaled to fill the budget (aspect preserved, 32-px aligned), saving you a separate Scale Image node and a lossy double-resize.
• negative_prompt - what to avoid. Optional but genuinely useful here.

Output is a single image.

Wiring it

Eric Qwen-Edit Loader (Qwen-Image-Edit-2511) → [Apply LoRA] → Eric Qwen-Edit Image → Preview/Save

Installing it

ComfyUI Manager → search "Eric Qwen-Edit", or:

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

Restart, then download the model - Qwen/Qwen-Image-Edit-2511, roughly 54 GB - and point the loader at its local directory. Use the pack's loader, not ComfyUI's stock "Load Diffusion Model": the stock nodes produce ComfyUI-internal wrappers that this pipeline can't consume. The pack's requirements pin diffusers>=0.34,<0.38 and transformers>=4.50,<5 - the transformers cap matters because 5.x dropped HybridCache that older diffusers imports; a startup import error is usually that.

VRAM expectations from the README: 24 GB for up to 2 MP, 48 GB for 6 MP, 96 GB for 16 MP. People do run it lower with sequential_offload on the loader - the trade is speed (roughly an order of magnitude slower).

Troubleshooting

• Edit looks identical to input - prompt too vague, or CFG too low. Qwen-Edit is instruction-driven; "make it better" does nothing.
• OOM - lower max_mp, or enable sequential_offload / attention_slicing on the loader.
• Slow - use the Lightning LoRA at 8 steps for iteration, save 50-step runs for finals.
• Whole image changes when you only wanted one region - that's Qwen-Edit's nature: it regenerates everything. For "leave everything else untouched", that's what the Inpaint node exists for.