Image editors fail in two opposite directions: they either change too little (the edit does not happen) or change too much (the whole image drifts). This release tests FireRed Image Edit 1.0 against Qwen 2511 across five edit types, with separate graphs for the anti-drift and free-range configurations.
The five tests
Background, lighting, prop, removal, recolour. That list is well-chosen because it ranges from structural (background) to subtle (recolour), and different editors fail differently across that range.
The graphs
| Workflow | Nodes |
|---|---|
| Single-image edit, anti-drift | 29 |
| Single-image edit, free-range | 28 |
| Two-image edit | 28 |
Nearly identical node counts — the difference between anti-drift and free-range is configuration, not structure. That is worth noting: the two modes are one graph with different sampling settings, not two different architectures.
The loader stack
| Node | Value |
|---|---|
UNETLoader |
FireRed-Image-Edit-1.0_fp8_e4m3fn.safetensors |
CLIPLoader |
qwen_2.5_vl_7b_fp8_scaled.safetensors, type qwen_image
|
VAELoader |
qwen_image_vae.safetensors |
EmptySD3LatentImage |
928 × 1664 |
LoraLoaderModelOnly |
Qwen-Image-Lightning-8steps-V2.0 bf16.safetensors at 1.0 |
The filename has a space in it
Qwen-Image-Lightning-8steps-V2.0 bf16.safetensors
There is a literal space between V2.0 and bf16. This is not a transcription artifact — it is the actual filename.
Two consequences:
- Shell and CLI paths break. Any command-line tool that splits on whitespace will fail unless the path is quoted.
- Download scripts with naive filename handling may save it without the space, and then the workflow's reference no longer matches.
If the model fails to load and you are certain it is present, check whether the filename on disk matches exactly, space included.
sd3 latent node on a Qwen model
EmptySD3LatentImage — the SD3 latent node — used with a Qwen encoder and Qwen VAE.
That is not an error; it is a latent-shape node reused across model families. But it means the node name does not tell you which model it belongs to, and anyone filtering nodes by "which model uses this" will get the wrong answer.
Anti-drift versus free-range
The naming says what each is for:
Anti-drift — the edit should be confined. The model is constrained to change only what the instruction names. Good for removal, recolour, and anything where the surrounding image must stay identical.
Free-range — the model is allowed to re-interpret the whole image. Good for background replacement and lighting changes, where a confined edit looks pasted.
Getting this wrong is the most common source of disappointing results: running a background replacement in anti-drift mode gives a subject floating on an unchanged background.
Why eight steps
Qwen-Image-Lightning-8steps-V2.0 at full weight. Eight steps is a compromise between four-step speed and the fidelity a careful edit needs.
For editing specifically, step count affects how much of the unchanged image survives. Fewer steps means more drift in the untouched regions. Eight is a reasonable floor for edits that must be conservative.
Where this breaks
- The space in the filename.
-
928 × 1664is fixed. Source images with other aspects get implicitly cropped or letterboxed. - FP8 quantisation on both UNET and encoder compounds detail loss.
- The anti-drift / free-range distinction is not documented in the files — you have to know which graph is which, and the node counts do not distinguish them.
- Eight steps is a floor, not a ceiling. For subtle edits, more steps helps.
How to use the comparison
Run all five edit types through both modes and score on: did the named edit happen, and did anything else change. The second question is the one that separates the modes, and it is the one people forget to check.
Workflows
Three graphs, published on RunningHub with invite code rh-v1111:
- FireRed Image Edit 1.0 single-image edit (anti-drift)
- FireRed Image Edit 1.0 single-image edit (free-range)
- FireRed Image Edit 1.0 two-image edit
Files
https://huggingface.co/AI-KSK/ksk-video-184-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)