你搞个面部修复，挂官方模型，然后微调模型lora 挂0.5，基本保持脸型一致，然后官方模型挂0.8lora，修复以后还原度很高。

Bro, your Chinese translation is actually really good — it already sounds pretty natural and conversational.

White was honestly one of the earlier people in our community to get disillusioned. Fuck the whole “begging and giving” mentality. We should all understand that full-parameter fine-tuning has been drifting further and further away from the community ever since F.1. There are very few people still willing to do full-parameter training while developing their own training methodology and model ecosystem around it. In the Chinese community, I know White, Easson, and only a handful of others.

That said, full-parameter fine-tuning of a DiT-based diffusion model can disrupt the model’s pretrained semantic and text-image alignment. Without the ability to explicitly control or customize the timestep / noise-level sampling distribution during training, even relatively conservative weight-space parameter injection can degrade high-frequency detail retention, reduce perceptual sharpness, and introduce reconstruction artifacts.

But sticking with this direction despite those constraints is still pretty damn difficult.

We’ve now managed to get PFM (Process-level / Process-supervised Flow Matching) running end-to-end — a new training paradigm based on process supervision — and we can actually train it under fairly constrained compute budgets. I still hope more people in the community will participate and push this research forward.

But I’ve completely given up on that old “beg for mercy and be grateful for everything” mentality. At some point, it started feeling less like a technical community and more like religious preaching.

I hope people can start evaluating these things more rationally. A full-parameter-trained model should generally be viewed as a Raw/Base checkpoint, not necessarily as a production-ready or universally compatible model. From that foundation, you still need multiple rounds of SFT / alignment refinement to recover and calibrate the desired behavior, followed by task-specific LoRA or adapter training for specialized domains and use cases.

So when a new training approach produces different characteristics, that doesn’t automatically mean the underlying model or the work behind it is “bad.” There are a lot of different points in the training and alignment pipeline, and each one comes with different trade-offs.

Hopefully this clears up some of your concerns. Just because things have changed doesn’t mean we should start hating on everything or dismissing everyone else’s work. Criticism is useful, but it should be based on understanding the training process and evaluating the actual failure modes, rather than simply comparing everything against the behavior of an older checkpoint.

As for the pricing, I see it as a pretty normal market mechanism. We can’t know exactly what someone else’s costs are, but we can know exactly what something is worth to us. So people should be free to price things according to their own needs, and then let actual market feedback push the supplier to adjust. Isn’t that basically how a market is supposed to work?