Eric Qwen Conditioned Edit

This is the node at the far end of Eric's conditioning system - the one that actually runs diffusion. Every other conditioning node (Encode, Interpolate, Blend, Direction Apply) is just building the input for this thing. If you've only ever used the pack's plain Eric Qwen-Edit Image node, think of this as the same edit, except you hand it a pre-computed embedding instead of a sentence.

Why would you bother? Because a conditioning is a number you can turn into a slider. Encode two conditionings - "neutral expression" and "smiling warmly" - interpolate between them at alpha 0.8, and this node renders that exact mix. No re-running the Qwen2.5-VL text encoder on every tweak, and no guessing whether the model "heard" your words the way you meant them.

How it works

Eric Qwen-Edit uses a dual conditioning path. The conditioning input you wire in here carries the VL semantic path - the token embeddings Qwen2.5-VL produced from your image plus instruction. The image input drives the VAE pixel path and sets the output resolution, which is why the tooltip reminds you it's usually the same image you encoded.

Under the hood it's the same real diffusers pipeline the whole pack is built on - FlowMatchEulerDiscreteScheduler, no Aura Flow Shift nodes, native resolution preservation up to your max_mp cap. The one thing to internalize: true CFG is off until you connect a negative_conditioning. If that input is empty, true_cfg_scale is ignored entirely. That trips people up, because the plain Edit Image node has CFG on by default.

The inputs that matter

• conditioning - your built embedding, straight from an Encode node or after passing through Interpolate / Direction Apply / Blend.
• negative_conditioning - optional, but it's the CFG switch. Connect it and true_cfg_scale goes live.
• true_cfg_scale - 1.5 is a gentle nudge, 3.0+ is strong. For direction-based edits the tooltip's advice is smart: start low (1.2–1.5), because the direction already steers and CFG on top of a direction is two steering wheels.
• max_mp - output ceiling (0.5–16). Lower is faster.
• upscale_to_max_mp - when on, small inputs get upscaled to fill the max_mp budget instead of you doing it with a Scale Image node first. Cheaper, because it avoids a lossy double-resize.

steps and seed work like every other node in the pack: 8 for a Lightning LoRA, 50 for the base model.

Wiring it up

Loader → Apply LoRA (PixelSmile) → Conditioned Edit
   ├─ Encode(image, "neutral expression") ──→ Interpolate ──┐
   └─ Encode(image, "smile warmly") ────────────→ alpha=0.8 ──┘

Installing it

Search "Eric Qwen-Edit" in ComfyUI Manager, or:

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

Restart ComfyUI. You need a Qwen-Image-Edit-2511 (or 2509) model directory - download it from Qwen/Qwen-Image-Edit-2511, roughly 54 GB. This pack runs the actual diffusers pipeline, so its loaders are mandatory - ComfyUI's built-in "Load Diffusion Model" wrappers won't work here. VRAM guidance from the README: 24 GB for up to 2 MP, 48 GB for 6 MP, 96 GB for 16 MP.

Troubleshooting

• No effect from CFG changes - negative_conditioning isn't connected. That's the switch.
• Output looks like the plain edit with no direction - check your Interpolate/Direction Apply actually changed conditioning; alpha 0 is a passthrough, by design.
• OOM on a big image - drop max_mp, or enable sequential_offload on the loader (it'll be slow, but it fits).