Studio Poche
The “Studio Pro” ComfyUI workflow: a phone photo becomes a magazine photo
- delivered master, from a 768×540 photo
- 3600×2400
- from photo to master, measured on an M1 Pro 16 GB
- 201 – 235 s
- stages, 26 nodes in a single workflow
- 7
- all local, no API
- 0 €
- The problem
- Image generators reinvent the car; plain development keeps it, but without the light of a real shoot.
- For whom
- Individuals, sellers and garages wanting a magazine-grade photo of their real car, without a studio or editing software.
- The outcome
- A 7-stage ComfyUI workflow: from a 768×540 phone photo to a 3600×2400 master in 201 to 235 s, car preserved.
Context
A text-to-image model reinvents the car: wheels change, bodywork warps. The first version of Studio Poche kept the car pixel-exact with a restrained look; Studio Pro adds the light of a real shoot.
Studio Pro gives the photo to the model both as a reference and as a starting point, then restores the car’s real detail by frequency separation: light comes from the render, fine detail from the photo.
What I built
- Lightroom-style development: robust white balance, auto exposure, highlights and shadows, denoising, sharpening.
- Generative relighting with Qwen-Image-2.1 locally: the photo is both the reference (Qwen3-VL encoder) and the starting latent (0.8 denoise), in 4 steps thanks to the Lightning LoRA.
- Photoshop-style subject fidelity: masking, aligning the original onto the render, then frequency separation inside the mask.
- ×4 upscale by tiled Real-ESRGAN with margins (SPUpscale node), then optics, film grading and grain; master, before/after sheet and manifest.
Architecture 7 blocks
- 01 · InputPhotoPhone photo, JPEG or PNG
- 02 · ToolDevelopmentWhite balance, exposure, denoising
- 03 · ModelQwen-Image-2.1Photo as reference and latent, 4 steps
- 04 · ToolFidelityMask, alignment, frequency separation
- 05 · ModelReal-ESRGAN ×4Tiles with margins, 3:2 format
- 06 · ToolOptics and gradingHalation, aberration, film look, grain
- 07 · OutputDeliveryMaster, before/after sheet, manifest
Pipeline 7 steps
- PhotoAny JPEG or PNG.
- DevelopThe “Lightroom” pass.
- RelightQwen-Image, photo as reference.
- FidelityThe real car’s detail.
- UpscaleReal-ESRGAN ×4, 3:2 format.
- Grade85 mm optics, film, grain.
- DeliverMaster, sheet, manifest.
Decisions the choice, the reason, the alternative ruled out
- D1Give the photo to the model as both reference and starting latent (0.8 denoise).
Without a reference, the car turned into another sedan; with it, the bodywork and geometry hold.
Ruled outA low denoise without reference, tried then dropped.
- D2Restore the subject’s real detail by frequency separation.
Light and colour come from the render, fine detail from the real photo: the car stays recognisable.
- D3Render below original size and leave high definition to Real-ESRGAN.
Enlarging before Qwen created a repeated texture pattern; the SPUpscale node also avoids the stock node’s stripes on MPS.
- D4One GPU render at a time.
Two MPS processes in parallel crashed the Metal driver.
Results
Measurements on 20 Oct 2026 (M1 Pro 16 GB)
| Example | Source | Qwen render | Master | Total time |
|---|---|---|---|---|
| Peugeot 308, real phone photo | 768×540 | 672×480, 4 steps | 3600×2400 | 201–228 s |
| Hypercar, photo degraded phone-style | 960×720 | 672×512, 4 steps | 3600×2400 | 230–235 s |
What iterating taught us
| Finding | Fix |
|---|---|
| Striped upscale on Apple MPS with the stock node | New SPUpscale node: tiles with margins (model correct, ×23 faster than CPU) |
| Repeated texture pattern if the photo is enlarged before Qwen | Render below original size, high definition comes from Real-ESRGAN |
| Without a reference, the car turns into another sedan | Photo as reference and 0.8 denoise |
| Two GPU processes in parallel crash Metal | One render at a time |
Deliverables 3
- Studio Pro workflow in editor and API versions, generated from a single spec.
- The code of the 9 Studio Poche nodes, installed as a ComfyUI pack.
- Two complete examples: intermediate stages, 3600×2400 masters, sheets and manifests.
Environment
Hardware and measurement setup
- Measured on a MacBook Pro M1 Pro 16 GB, one GPU render at a time.
- Generative part under the Qwen Research licence (non-commercial).
Gallery 9
What this project means for you
The same know-how, on your problem.
Your photos, magazine version
A phone photo becomes a presentable visual, and your product stays exactly yours.
A workflow your team opens
Seven readable groups in ComfyUI, every setting visible and editable.
Nothing leaves your premises
100% local processing, nothing sent to an online service.