JoyAI-Image-Edit vs Qwen 2511: Progressive Editing Is the Real Gap
Give both models the same two-step instruction — put the apple in the red box, then remove the box — and only one of them does both. JoyAI applies what it understood continuously during diffusion; 2511 tends to execute a
Give both models the same two-step instruction — put the apple in the red box, then remove the box — and only one of them does both. JoyAI applies what it understood continuously during diffusion; 2511 tends to execute a single clause and stop. Eleven head-to-head tests show where that matters and where it does not.
Evidence note: two workflow JSON files ship with this episode and were parsed directly; node counts, UNET/CLIP/VAE filenames, LoRA strengths and sampler settings below come from those files.
When this is useful
- Two-step edits that must both land, such as move an object then remove the container.
- Camera control: yaw, pitch and zoom specified numerically in the prompt.
- Viewpoint changes on a product or a piece of furniture.
- Single-pass clothing swaps written as remove-then-add.
How this differs from the previous build
The difference is in how the instruction is consumed, not in raw fidelity.
- Progressive application. JoyAI keeps re-applying the understood condition through diffusion; 2511 generally executes one clause of a multi-part prompt.
- Camera control is a first-class task. The author's yaw/pitch/zoom prompts move JoyAI and barely move 2511.
- Single-image only. The non-Plus build does not do multi-image editing, where 2511 wins almost by default.
- NSFW tolerance enables one-pass swaps. Writing remove-then-add as one prompt works here; with 2511 the same job takes two separate edits.
Running it
- Load the image and set the processing size. About two million pixels is the author's working figure, with no clear degradation.
- Write the instruction. For compound tasks, write both steps explicitly and in order.
- For camera work, state the movement numerically — yaw, pitch, zoom.
- Run the single-image graph for JoyAI; use the comparison graph when you want both models on the same input.
- For clothing swaps, write it as remove the existing garment, then add the new one, in a single prompt.
- Compare outputs side by side; the comparison graph is built to emit one result per model.
Parameters, and where the numbers come from
| Item | Value in the shipped graphs |
|---|---|
| JoyAI sampler |
JoyAI_Image_SM_KSampler, steps 20, guidance 7.5 |
| Comparison UNETs |
qwen_image_edit_2511_bf16.safetensors, z_image_turbo_bf16.safetensors
|
| Comparison text encoders |
qwen_2.5_vl_7b_fp8_scaled.safetensors (qwen_image), qwen_3_4b.safetensors (qwen_image) |
| Comparison VAEs |
qwen_image_vae.safetensors, ae.safetensors
|
| LoRA, 2511 accelerator |
Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors at 1.0 |
| LoRA, Z-Image |
zit_fdpo_v1.safetensors at 0.8, lenovo__z.safetensors at 0.2 |
| Comparison samplers | KSampler steps 4 / CFG 1 / euler / simple; KSampler steps 9 / CFG 1 / euler / simple
|
The two KSampler entries are the accelerated comparison lanes, not the JoyAI path — JoyAI runs through its own sampler node with guidance 7.5.
Honest limits
- JoyAI broke on at least one camera-control test, so the camera advantage is not clean.
- 2511 keeps real strengths: industrial materials and overall commercial polish were better in the material-conversion test.
- Multi-image editing is a straight loss for the non-Plus JoyAI build.
- Generation speed is the author's main reservation about replacing 2511 outright.
- The eleven comparisons are visual judgements without scores, and several are described as ties.
Episode resources
RunningHub workflows
- JoyAI Image single-image edit — https://www.runninghub.ai/post/2043308514138918913/?inviteCode=rh-v1111
- JoyAI Image vs 2511 comparison — https://www.runninghub.ai/post/2043308481817616386/?inviteCode=rh-v1111
- Video tutorial (YouTube): https://www.youtube.com/watch?v=uv7JK5-X0jc
- Chinese original (Bilibili): https://www.bilibili.com/video/BV1MEQbBsE5z
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.