Audit and harden a production image & video generation pipeline
Can you explain why a hand comes out distorted, or why a face drifts a few seconds into a video, and prove it on a fixed seed before you fix it? Then this role is for you.
About us
Connct is an AI companion app: photorealistic characters that users chat with, and that send them photos and videos of themselves.
We have two non-negotiable standards:
- every character must stay instantly recognizable from one piece of media to the next;
- every piece of media must look believable at first glance.
We produce this media with our own in-house generation factory ("the Lab"), running on on-demand GPUs. Today, every single image is still reviewed by hand. We're looking for someone who truly masters this field to audit the entire pipeline, tell us, with evidence, what needs to change, and then implement it.
The mission
Phase 1: full audit of the generation pipeline
- Image quality: anatomy and proportions (hands, limbs, body), consistency of held objects, character identity stability across images.
- Video quality: natural motion from the very first second, the right limbs moving, no face drift, output faithful to the requested action.
- Prompt system: is the way we write, assemble and rewrite prompts right for these models? What belongs in the prompt, in a LoRA, in a ControlNet, in an inpainting pass?
- Models and LoRAs: model selection, weights, whether to train specialized LoRAs (anatomy, motion), and how.
- Automation: produce full weekly series for 18 characters (and more very soon) with no human intervention and zero waste.
- Infrastructure and GPU costs: GPU selection, pod lifecycle, reliability.
Phase 2: implementation of the approved fixes
Deliverable: a prioritized report in the format problem → proven root cause → fix → expected gain, followed by implementation of the approved fixes.
Our stack
Image
- ComfyUI driven exclusively via API (JSON workflows, no manual UI use)
- Z-Image-Turbo (bf16), 2 sampling passes, SeedVR2 upscaling up to 4K
- 18 in-house trained character LoRAs, stacked with style and realism LoRAs
- FaceDetailer / Impact-Pack, rgthree
- LoRA training with ai-toolkit (datasets and captions generated in-house)
Video
- LTX-2.3 image-to-video (22B distilled, x2 spatial upscaler, Gemma 3 12B text encoder, fp8)
- Tiled sampler, face anti-drift anchors, segment chaining for long-form clips
Prompts
- Prompts assembled automatically from building blocks (Python), gated by a blocking lint
Infrastructure and orchestration
- RunPod: on-demand GPU pods (RTX PRO 6000, H100, RTX 5090…), shared network volumes
- In-house Lab built in TypeScript / Node (Hono API, BullMQ, SQLite), Next.js review UI
- S3 storage with catalog tags, consumed by the app (Laravel 12 / PostgreSQL)
What we're looking for
Must-haves
- Real, recent experience running ComfyUI in production (API workflows, custom nodes, graph debugging), not just occasional creative use
- Strong command of recent diffusion models (Z-Image, Flux, SDXL) and image-to-video (LTX, Wan, Hunyuan or equivalent)
- LoRA training (image, ideally video/motion): datasets, captioning, hyperparameters, checkpoint selection
- Inpainting, ControlNet / pose, detailers, upscalers
- Methodical rigor: one variable at a time, fixed-seed comparisons, proof before conclusions
Nice-to-haves
- Automated image QA (defect detection, scoring)
- Cloud GPU experience (RunPod or equivalent), VRAM / fp8 optimization
- Python and/or TypeScript to work directly in the pipeline
- Consistent character generation at scale
What we offer
- A real pipeline already in production, with growing volumes and genuinely hard technical problems
- Direct access to the tech team and founders, fast decisions
- Full autonomy over your methods
- [Dedicated GPU budget for testing]
- [Potential long-term collaboration beyond this mission]
Terms
- Short term contract first
Please send:
- Your resume or online profile (LinkedIn, GitHub, Hugging Face, Civitai…)
- Two or three concrete examples of production work: a ComfyUI workflow, a trained LoRA, an image-to-video pipeline…
- A few lines on a quality issue you solved (anatomy, identity drift, motion) and how you proved your fix was the right one.
Applications that include concrete examples will be reviewed first.
Pay: AED250.00 - AED400.00 per hour
Expected hours: 1.0 – 10.0 per week
Work Location: Remote