Chinese teams are at the front of generative video, which is the hardest thing currently being done with diffusion.
Image generation has become widely understood. Video has not, and it is where diffusion engineering stops being tractable: temporal consistency across frames, plausible motion, and compute economics that punish every wasted step in both training and inference.
Kuaishou’s Kling, MiniMax’s Hailuo and Alibaba’s Wan family put Chinese engineering at or near the front of text-to-video, and Tencent and others have released open image and video models that the wider community now builds on. Engineers from those teams have worked on problems most image-generation specialists have never had to face.
A great many people can adapt an existing checkpoint with a LoRA. Far fewer can tell you why your generations degrade at higher resolutions, or which sampler and step count is costing you quality rather than saving you money.
The screening question that separates them reliably: ask how they would halve inference cost without a visible drop in output quality. Strong answers move through distillation and reduced-step sampling, resolution strategy, caching where prompts repeat, and batching — and they will insist on a human evaluation set before and after, because automated metrics in this field correlate poorly with what people actually think looks good. Weak answers reach for a smaller model and stop.
This is the part most teams underestimate. There is no equivalent of an accuracy score: whether a generation is good is a judgement, and the only reliable instrument is a consistent panel of people looking at paired outputs.
Ask candidates how they have run that. You are listening for a fixed prompt set held back from training, blind pairwise comparison, more than one rater, and some account of rater disagreement. Engineers from the large Chinese media platforms have usually been close to exactly this machinery, because those companies ship generative features to consumers who complain immediately.
Because so much of the significant work has been published as open weights, you can check a candidate’s claims before you speak to them. Ask which release they contributed to, then read the model card, the repository and the paper.
The usual caution applies about long author lists on large releases — establish what this person did. Our note on verifying Chinese engineering credentials covers the cases where the contribution is real but harder to attribute.
Worth separating before you brief, because the skill sets diverge more than the job titles suggest. Image work is largely about control and fidelity: conditioning, LoRAs, inpainting, getting a consistent character or product across a set. Video adds the entire problem of time — motion coherence, frame-to-frame stability, and training costs that are an order of magnitude higher.
An engineer strong on image generation is not automatically able to help with video, and hiring for video when you need image work means paying for experience you will not use. If you genuinely need both, hire the video engineer: the reverse rarely works.
First, training data provenance. Generative media sits in a live legal argument about what may be trained on, and the answer differs between the UK, the US and China. Decide what your model may see, write it into the contract, and do not leave it to a contractor’s judgement. Our note on IP assignment with Chinese contractors covers how to treat the data and the resulting weights, which are separate questions.
Second, compute. This is expensive work, and whether the engineer uses your hardware or their own changes the rate, the contract and the IP position. Settle it at the briefing stage rather than discovering it in the first invoice.
Day rates, invoiced by our UK company in sterling. Reviewed September 2026. The figure depends on seniority and on how much overlap with your working day you need.
How the engagement works — they work directly for your team, you brief and manage them, and the money runs through one UK invoice.
Our guide to hiring diffusion model engineers in the UK covers what the domestic market pays, how to screen, and the interview questions that actually discriminate.
Describe the work and the overlap hours you need, and we will come back with profiles and a rate. No retainer, and nothing to sign before you have seen candidates.
Start hiring