Watch the full video first if you want to understand how this FLUX.2 image restoration workflow works in practice. The video shows how a source image can be loaded, preserved as a reference latent, and regenerated through a controlled FLUX.2 image-to-image route to improve clarity, texture, edge quality, and overall high-resolution detail.
This ComfyUI workflow is designed for FLUX.2 image restoration and high-resolution image repair. Its main purpose is not to redesign the input image. Instead, it uses the loaded source image as the exact reference, then improves the image quality while preserving the original subject, identity, composition, pose, perspective, proportions, colors, lighting, background layout, typography, and important visual details.
The workflow is built around a FLUX.2 distilled 8-step FP8 mixed model route. It uses flux2_distilled_8step_fp8mixed.safetensors as the main diffusion model, a Mistral 3 Small Flux2 text encoder, and flux2-vae.safetensors for latent encoding and decoding. The generation path includes BasicGuider, FluxGuidance, Flux2Scheduler, KSamplerSelect, RandomNoise, SamplerCustomAdvanced, VAEDecode, and SaveImage.
The first important part of the workflow is the source image preparation section. LoadImage imports the original image. image_scale_pixel_v2 prepares the image for processing while keeping the visual source aligned. GetImageSize reads the source image width and height, and the Flux2Scheduler follows that image size during sampling. This means the workflow is not simply generating on a blank canvas; it is directly tied to the source image dimensions.
The second important part is the reference latent route. The prepared source image is encoded through VAEEncode into latent space. That latent is then used by ReferenceLatent and sent into the conditioning route. This gives the model a stronger reference anchor during restoration. The goal is to preserve the original structure while allowing FLUX.2 to rebuild cleaner details, sharper edges, better high-frequency texture, and more polished output quality.
The prompt section is intentionally simple. PiDTextPrompt is set to “high resolution,” and the CLIPTextEncode node uses a structured JSON-style repair instruction. The instruction tells the model to preserve the source image exactly while improving clarity, material detail, edge definition, and clean high-resolution quality. It also explicitly tells the workflow not to invent new objects, change style, crop the image, alter the camera angle, or redesign the image.
The sampling section uses FLUX.2’s 8-step distilled route. FluxGuidance controls the strength of the conditioning, while BasicGuider sends the model and conditioning into SamplerCustomAdvanced. The sampler uses the source image latent, noise, guider, sampler selection, and scheduler sigmas to generate the restored result. Finally, VAEDecode converts the latent back into an image, and SaveImage exports the result.
Compared with ordinary upscale workflows, this workflow is more useful when the goal is visual restoration rather than pure size enlargement. It is suitable for blurry image repair, AI image cleanup, compression softness removal, detail enhancement, edge recovery, texture rebuilding, poster cleanup, character image refinement, product image restoration, and high-resolution preview generation.
Main features:
FLUX.2 image restoration workflow
Source image preservation route
Image-to-image latent reconstruction
FLUX.2 distilled 8-step FP8 mixed model
Mistral 3 Small Flux2 text encoder
Flux2 VAE encoding and decoding
LoadImage source image input
image_scale_pixel_v2 image preparation
GetImageSize source-size tracking
VAEEncode source latent creation
ReferenceLatent conditioning support
PiDTextPrompt high-resolution instruction
JSON-style image repair prompt
FluxGuidance control
Flux2Scheduler follows source image size
SamplerCustomAdvanced generation route
SaveImage final output
Suggested workflow:
Prepare one image that needs restoration or quality enhancement. The workflow works best when the original image already has a clear subject and readable composition. Load the image into the workflow, keep the prompt focused on restoration, and avoid adding new creative concepts unless you intentionally want the image to change. Start with the default 8-step route first. If the result changes the image too much, reduce creative wording and strengthen preservation language. If the result is still too soft, increase the clarity, high-frequency texture, edge definition, and material detail wording. Use this workflow when you want to clean and enhance an image without turning it into a completely different picture.
⚙️ RunningHub Workflow
Try the workflow online right now — no installation required.
👉 Workflow: https://www.runninghub.ai/post/2064329697353359361?inviteCode=rh-v1111
If the results meet your expectations, you can later deploy it locally for customization.
🎁 Fan Benefits: Register to get 1000 points + daily login 100 points — enjoy 4090 performance and 48 GB super power!
📺 Bilibili Updates (Mainland China & Asia-Pacific)
If you’re in the Asia-Pacific region, you can watch the video below to see the workflow demonstration and creative breakdown.
📺 Bilibili Video: https://www.bilibili.com/video/BV1RtEQ6SEsS/
☕ Support Me on Ko-fi
If you find my content helpful and want to support future creations, you can buy me a coffee ☕.
Every bit of support helps me keep creating — just like a spark that can ignite a blazing flame.
👉 Ko-fi: https://ko-fi.com/aiksk
💼 Business Contact
For collaboration or inquiries, please contact aiksk95 on WeChat.
⚙️打开下方链接即可在线体验，无需安装。
👉 工作流： https://www.runninghub.ai/post/2064329697353359361?inviteCode=rh-v1111
如果觉得效果理想，你也可以在本地进行自定义部署。
🎁 粉丝福利： 注册即送 1000 积分，每日登录 100 积分，畅玩 4090 体验 48 G 超级性能！
📺 Bilibili 更新（中国大陆及南亚太地区）
如果你在中国大陆或南亚太地区，可以通过下方视频查看该工作流的实测效果与构思讲解。
📺 B站视频： https://www.bilibili.com/video/BV1RtEQ6SEsS/
我会在 夸克网盘 持续更新模型资源：
👉 https://pan.quark.cn/s/20c6f6f8d87b
这些资源主要面向本地用户，方便进行创作与学习。