Qwen-Image-Edit-2511 — FP8 variants (img2img, RTX 5090)
Timings & VRAM
| Variant | CFG | Steps | Load | Avg gen / img | Peak VRAM | Success |
|---|---|---|---|---|---|---|
| base (non-distilled) | 4.0 | 50 | 56.1s | 81.8s | 22.28 GB | 9 / 9 |
| 8-step distilled | 1.0 | 8 | 59.3s | 8.6s | 22.27 GB | 9 / 9 |
| 4-step distilled | 1.0 | 4 | 50.7s | 5.5s | 22.27 GB | 9 / 9 |
lightx2v fp8 DiT (per-channel W8A8 via ISK fp8_dense_gemm_sm120) + our fp8 Qwen2.5-VL-7B text encoder (196 decoder Linears; vision tower bf16). DiT self-attention via SageAttention (int8-QK / fp8-PV; ~3× vs torch FlashAttention). Same seed (0) and resolution (1024×1024) across variants. Per-variant — base (non-distilled): 50 steps, cfg 4.0; 8-step distilled: 8 steps, cfg 1.0; 4-step distilled: 4 steps, cfg 1.0. Drag the slider to compare A vs B.
example #01
Change the person to a standing position, bending over to hold the dog's front paws.
example #02
Replace the words 'HEALTH INSURANCE' on the letter blocks with 'Tomorrow will be better'.
example #04
Replace the polka-dot shirt with a light blue shirt.
example #05
Change the background to Antarctica.
example #06
Generate a cartoon profile picture of the person.
example #07
Remove the hair from the plate.
example #08
Transform the character in the image into an anime style, and add the text: "Accelerate image generation and editing with Lightx2V Qwen-Image-Lightning".
example #13
Make the girl from Image 1 wear the black dress from Image 2 and sit in the pose from Image 3.
example #14
Make the girl from Image 1 wear the necklace from Image 2 and carry the bag from Image 3 on her left shoulder.
