Introduction

You know the shot: perfect lighting, perfect pose, and then a melted logo on the hoodie or a hand with one finger too many. ComfyUI Detail Fixer is a new open-source node pack from Somarawad, released at 17:33 UTC on October 31 (01:33 MYT, November 1). It fixes exactly those spots and leaves everything else pixel-identical.

It's MIT licensed and needs no extra Python packages. It also ships a ready example workflow. You can run it fully locally on a 10 GB GPU (SAM 3 plus a Qwen Image Edit 2509 GGUF), or send any single fix to Nano Banana 2 or GPT Image 2 through ComfyUI's built-in partner API nodes.

What Shipped

  • Detail Detect: finds targets with SAM 3 from a word or two, such as logo + text, face, hand or jeans. It can also use the free Gemini API or manual boxes. Touching boxes merge, so a brand name and its tagline get fixed together.
  • Detail Crop: cuts one upscaled, context-padded crop per detection. It outputs a box mask, plus SAM 3's outline as a shape mask when you connect it.
  • Detail Prepare Logo: erases the old print while keeping the fabric, then lays a faint guide of your real logo (around 0.6 opacity) so the model redraws it in the right place.
  • Detail Paste Back: blends each regenerated crop back along the object's outline, with optional colour and contrast matching for faces and skin.
  • Detail Pick Model: a per-slot switch with skip, local (Qwen), api 1 (Nano Banana 2) and api 2 (GPT Image 2). Unselected branches don't run, so API credits are only spent when you pick an API.
  • Example graph: examples/detail_fixer.json chains Logo → Face → Custom slots. It upscales the photo 2× first with 4x-UltraSharp at half scale, so small details get enough pixels.

How to Use It

cd ComfyUI/custom_nodes
git clone https://github.com/Somarawad/ComfyUI-DetailFixer

Restart ComfyUI and drag in examples/detail_fixer.json. For the logo slot, load your logo twice: once as the reference image and once as a transparent PNG. Then queue.

Models for the local path:

  • sam3.1_multiplex_fp16.safetensors from Comfy-Org/sam3.1 in models/checkpoints
  • Qwen-Image-Edit-2509-Q3_K_S.gguf in models/unet/gguf (loaded with the gguf nodes)
  • Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors in models/loras
  • qwen_2.5_vl_7b_fp8_scaled.safetensors in models/text_encoders
  • qwen_image_vae.safetensors in models/vae
  • 4x-UltraSharp.pth in models/upscale_models

The API branches use your Comfy.org credits. Sign in under Settings.

Suggested Strengths

  • Logo: find logo + text, denoise 0.7–0.8, with erase and guide on and your logo as the reference
  • Face: find face, denoise 0.35–0.45, with outline blend and colour match, and an optional face reference
  • Hands: find hand, denoise 0.4–0.5, with outline blend and colour match
  • Clothing: find jeans or jacket, denoise 0.4–0.6, and describe the fix, for example "remove the extra pocket"

Tips from the author: if the edit model ignores your instruction, raise context in Detail Crop, because very tight crops make edit models copy their input. On a dark garment, use a light version of your logo for the guide, and vice versa. Only use logos and likenesses you have the rights to.

Good to Know

  • Very new: the repo had 0 stars when we checked. It does include unit tests and a before/after sample.
  • Local vs API: the local path uses an aggressive Q3 GGUF to fit 10 GB. Expect the API slots to give crisper typography on hard logos, at a credit cost.

Conclusion

Detail Fixer turns "almost perfect" into "ship it" without re-rolling the whole image. Point SAM 3 at the problem, regenerate just that patch and blend it back invisibly. It's a tidy, well-documented graph that's worth keeping in your finishing toolkit.

Original Source

Fix the details, and happy creating!

—Aurelia ♡