ComfyUI / Flux / Z-Image Beauty Filter — Visual Tuning Specialist
Budget: $30 – $250 USD
Project: ComfyUI / Flux / Z-Image Beauty Filter — Visual Tuning Specialist
I'm running a production AI beauty retouching web service (photographerfilter.com). The full pipeline is built and deployed — what I need is an experienced ComfyUI / Flux / LoRA specialist to dial in the final visual quality. This is a tuning job, not infrastructure work.
Pipeline (already running on Modal / A100):
Stage 1: Flux 2 Klein 9B + custom Beauty LoRA + ultra_real LoRA
Stage 2: Z-Image Turbo + Kook LoRA + ControlNet tile (masked, face/body)
MediaPipe FaceMesh eye-preservation blend
Custom LAB skin-tone correction (Python)
Stage 3: SeedVR2 GGUF upscale to ~2048px
Side-by-side composite output
Visual problems I need solved:
Skin texture too fine / too smooth — natural pore patterns are removed, reads as plastic / airbrushed.
Eyes not preserved well — still glassy, catch lights (iris reflections) look unrealistic.
Faces over-lit, especially on dark-skinned models — washed out, ashy, not properly toned.
The beautifying side (blemish removal, skin clarity, warm tones, sculpting) is already excellent. The work is bringing realism back without losing it.
You'll receive:
A full handoff document with everything learned to date (~500 pairs of trial/error condensed)
index.html, beauty_phase_a.py, skin_tone_correct.py for context
~500 before/after pairs (original photo + ideal manual retouch) as your aesthetic target
A tester account on the live site with 500 credits (one credit = one full pipeline run on real GPU)
10 numeric dials and 4 prompt fields exposed through an admin panel — tune any of them per job
Workflow:
Express interest. Sign in at photographerfilter.com with a Google account. Send me the email you used.
I promote your account to "tester" — panel appears automatically, 500 credits auto-granted.
You iterate through the panel. Each job runs the live pipeline with your current settings.
When you're happy with a result, save the configuration with a name.
I load your saved configs from my admin side and test them against my own evaluation set.
Payment is released once I review your saved config(s) against my own test images and love the results.
Critical heads-up: The text encoders are Qwen 3 (8B / 4B), which is brutally literal compared to CLIP/T5. "Preserve skin texture" → dragon scale. "8K resolution" → painfully crispy. "Subsurface scattering" → glassy eyes. A huge amount of what's already known is in the handoff so you don't redo it.
Required experience:
ComfyUI workflows with Flux family + Z-Image Turbo
Custom LoRA stacking and weight tuning for portraits
Prompt engineering on modern LLM-based text encoders (Qwen / Gemma / T5-XXL)
LAB color-space skin tone work
Patience to test across varied skin tones, lighting, and crops
Out of scope:
Modal / Cloudflare / deployment / backend
Retraining LoRAs
New pipeline stages
Frontend / billing / auth changes
Budget: Fixed price on successful delivery — discuss with me. I'm prioritizing the right person over the cheapest bid.
To apply: Tell me briefly about a Flux or Z-Image project you've tuned for portrait work and what changed when you got it right. Then sign in at photographerfilter.com and send me the email.
I'm running a production AI beauty retouching web service (photographerfilter.com). The full pipeline is built and deployed — what I need is an experienced ComfyUI / Flux / LoRA specialist to dial in the final visual quality. This is a tuning job, not infrastructure work.
Pipeline (already running on Modal / A100):
Stage 1: Flux 2 Klein 9B + custom Beauty LoRA + ultra_real LoRA
Stage 2: Z-Image Turbo + Kook LoRA + ControlNet tile (masked, face/body)
MediaPipe FaceMesh eye-preservation blend
Custom LAB skin-tone correction (Python)
Stage 3: SeedVR2 GGUF upscale to ~2048px
Side-by-side composite output
Visual problems I need solved:
Skin texture too fine / too smooth — natural pore patterns are removed, reads as plastic / airbrushed.
Eyes not preserved well — still glassy, catch lights (iris reflections) look unrealistic.
Faces over-lit, especially on dark-skinned models — washed out, ashy, not properly toned.
The beautifying side (blemish removal, skin clarity, warm tones, sculpting) is already excellent. The work is bringing realism back without losing it.
You'll receive:
A full handoff document with everything learned to date (~500 pairs of trial/error condensed)
index.html, beauty_phase_a.py, skin_tone_correct.py for context
~500 before/after pairs (original photo + ideal manual retouch) as your aesthetic target
A tester account on the live site with 500 credits (one credit = one full pipeline run on real GPU)
10 numeric dials and 4 prompt fields exposed through an admin panel — tune any of them per job
Workflow:
Express interest. Sign in at photographerfilter.com with a Google account. Send me the email you used.
I promote your account to "tester" — panel appears automatically, 500 credits auto-granted.
You iterate through the panel. Each job runs the live pipeline with your current settings.
When you're happy with a result, save the configuration with a name.
I load your saved configs from my admin side and test them against my own evaluation set.
Payment is released once I review your saved config(s) against my own test images and love the results.
Critical heads-up: The text encoders are Qwen 3 (8B / 4B), which is brutally literal compared to CLIP/T5. "Preserve skin texture" → dragon scale. "8K resolution" → painfully crispy. "Subsurface scattering" → glassy eyes. A huge amount of what's already known is in the handoff so you don't redo it.
Required experience:
ComfyUI workflows with Flux family + Z-Image Turbo
Custom LoRA stacking and weight tuning for portraits
Prompt engineering on modern LLM-based text encoders (Qwen / Gemma / T5-XXL)
LAB color-space skin tone work
Patience to test across varied skin tones, lighting, and crops
Out of scope:
Modal / Cloudflare / deployment / backend
Retraining LoRAs
New pipeline stages
Frontend / billing / auth changes
Budget: Fixed price on successful delivery — discuss with me. I'm prioritizing the right person over the cheapest bid.
To apply: Tell me briefly about a Flux or Z-Image project you've tuned for portrait work and what changed when you got it right. Then sign in at photographerfilter.com and send me the email.