DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

JoyAI Image Edit vs Qwen 2511: Where the Gap Actually Shows Up

Two image editors, same task, similar architecture — and the difference turns out to be spatial understanding and stepwise prompt execution rather than raw quality. That is a useful result because those are testable properties, not taste.

The comparison graph

The 51-node compare workflow loads the Qwen side explicitly:

Node Value
UNETLoader qwen_image_edit_2511_bf16.safetensors
CLIPLoader qwen_2.5_vl_7b_fp8_scaled.safetensors, type qwen_image
VAELoader qwen_image_vae.safetensors
LoraLoaderModelOnly Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors at 1.0
EmptySD3LatentImage 928 × 1664

The 13-node single-edit graph is the JoyAI side.

The resolution is the interesting detail

928×1664 is 1.54 megapixels, and it is not a round number. Round numbers (1024×1024, 1280×720) are chosen for convenience. 928×1664 looks like it was chosen for a reason — most likely an aspect that suits the edit tasks being compared.

Aspect matters more in editing than in generation. When you edit an existing image, the output aspect has to match the input aspect, otherwise the model is either cropping or letterboxing implicitly. A latent size that matches your source aspect avoids that entirely.

If your source is portrait and close to 9:16, 928×1664 (which is 0.558, just above 9:16's 0.5625) is nearly a direct match.

The Lightning LoRA at full weight

Qwen-Image-Edit-2511-Lightning-4steps-V1.0 at weight 1.0 means four steps.

Four-step distillation on an editor is a different proposition than on a generator. An editor has to preserve the parts of the image it is not changing; a distilled student trained to work in four steps may not have the budget to be conservative about preservation.

That is the trade: speed versus how much of the untouched image survives unchanged. If you see unexpected changes outside your edit region, step count and LoRA weight are the levers.

What the comparison is actually measuring

The claim is that the gap shows up in spatial understanding and stepwise instruction execution. Both are worth defining, because they determine how you should test:

Spatial understanding — "move the cup from the left of the plate to behind it". This requires the model to know where things are relative to each other, not just what they are. A model with weak spatial grounding will move the cup somewhere, or move the plate instead.

Stepwise instruction execution — "remove the hat, then add a scarf, then change the background to evening". Three sequential edits. A model that handles these as a bag of keywords will do one or two; a model that handles them as a sequence will do all three without the operations interfering.

These are checkable with a fixed set of prompts, and a comparison built on them is more informative than one built on "which looks better".

Where a comparison like this is weak

  1. Two models, one prompt set. Conclusions are specific to the prompts chosen. A different set can reverse the ranking.
  2. The 13-node JoyAI graph is minimal. It has almost no supporting machinery, so a fair comparison would need to match operational overhead too.
  3. Four-step distillation favours whoever is better at short-budget preservation, which is not necessarily the better editor at full step count.
  4. fp8 on the text encoder (qwen_2.5_vl_7b_fp8_scaled) is a quantisation the comparison inherits; it may affect prompt understanding more than image quality.

How to run this comparison yourself

Fix one source image and one three-step instruction. Render with both models at matched seed. Then score on: did all three steps happen, did they happen in order, and did anything outside the edit regions change. That gives you a result you can act on.

Workflows

Two graphs, published on RunningHub with invite code rh-v1111:

  • JoyAI Image single-image edit
  • JoyAI Image compared against Qwen 2511

Files

https://huggingface.co/AI-KSK/ksk-video-220-comfyui-workflows

Notes

  • Workflow files only. Model weights are not included; each keeps its own licence.
  • Tutorial exports, not a product.
  • Node counts and loader values quoted above are read from the published JSON files.
  • Independent AI-KSK release. Not affiliated with the model authors or RunningHub.

Optional API platform - 88API (affiliate link)

Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.

https://88api.ai/sign-up?aff=usx3

Disclosure: I may earn a commission from eligible purchases after you register through this referral link.

Top comments (0)