replay

Diffusion fashions have lately emerged because the de facto commonplace for producing complicated, high-dimensional outputs. It’s possible you’ll know them for his or her capacity to provide beautiful AI artwork and hyper-realistic artificial photos, however they’ve additionally discovered success in different functions resembling drug design and steady management. The important thing thought behind diffusion fashions is to iteratively rework random noise right into a pattern, resembling a picture or protein construction. That is sometimes motivated as a most probability estimation drawback, the place the mannequin is educated to generate samples that match the coaching information as carefully as doable.

Nonetheless, most use circumstances of diffusion fashions usually are not instantly involved with matching the coaching information, however as a substitute with a downstream goal. We don’t simply need a picture that appears like current photos, however one which has a selected sort of look; we don’t simply desire a drug molecule that’s bodily believable, however one that’s as efficient as doable. On this submit, we present how diffusion fashions may be educated on these downstream targets instantly utilizing reinforcement studying (RL). To do that, we finetune Secure Diffusion on a wide range of targets, together with picture compressibility, human-perceived aesthetic high quality, and prompt-image alignment. The final of those targets makes use of suggestions from a big vision-language mannequin to enhance the mannequin’s efficiency on uncommon prompts, demonstrating how highly effective AI fashions can be utilized to enhance one another with none people within the loop.

A diagram illustrating the prompt-image alignment goal. It makes use of LLaVA, a big vision-language mannequin, to guage generated photos.

Denoising Diffusion Coverage Optimization

When turning diffusion into an RL drawback, we make solely essentially the most primary assumption: given a pattern (e.g. a picture), we now have entry to a reward perform that we are able to consider to inform us how “good” that pattern is. Our aim is for the diffusion mannequin to generate samples that maximize this reward perform.

Diffusion fashions are sometimes educated utilizing a loss perform derived from most probability estimation (MLE), that means they’re inspired to generate samples that make the coaching information look extra doubtless. Within the RL setting, we now not have coaching information, solely samples from the diffusion mannequin and their related rewards. A technique we are able to nonetheless use the identical MLE-motivated loss perform is by treating the samples as coaching information and incorporating the rewards by weighting the loss for every pattern by its reward. This offers us an algorithm that we name reward-weighted regression (RWR), after current algorithms from RL literature.

Nonetheless, there are a couple of issues with this method. One is that RWR just isn’t a very actual algorithm — it maximizes the reward solely roughly (see Nair et. al., Appendix A). The MLE-inspired loss for diffusion can be not actual and is as a substitute derived utilizing a variational certain on the true probability of every pattern. Which means RWR maximizes the reward by two ranges of approximation, which we discover considerably hurts its efficiency.

We consider two variants of DDPO and two variants of RWR on three reward features and discover that DDPO constantly achieves the very best efficiency.

The important thing perception of our algorithm, which we name denoising diffusion coverage optimization (DDPO), is that we are able to higher maximize the reward of the ultimate pattern if we take note of your complete sequence of denoising steps that received us there. To do that, we reframe the diffusion course of as a multi-step Markov determination course of (MDP). In MDP terminology: every denoising step is an motion, and the agent solely will get a reward on the ultimate step of every denoising trajectory when the ultimate pattern is produced. This framework permits us to use many highly effective algorithms from RL literature which can be designed particularly for multi-step MDPs. As a substitute of utilizing the approximate probability of the ultimate pattern, these algorithms use the precise probability of every denoising step, which is extraordinarily simple to compute.

We selected to use coverage gradient algorithms on account of their ease of implementation and previous success in language mannequin finetuning. This led to 2 variants of DDPO: DDPO_SF, which makes use of the straightforward rating perform estimator of the coverage gradient also called REINFORCE; and DDPO_IS, which makes use of a extra highly effective significance sampled estimator. DDPO_IS is our best-performing algorithm and its implementation carefully follows that of proximal coverage optimization (PPO).

Finetuning Secure Diffusion Utilizing DDPO

For our essential outcomes, we finetune Secure Diffusion v1-4 utilizing DDPO_IS. We now have 4 duties, every outlined by a distinct reward perform:

• Compressibility: How simple is the picture to compress utilizing the JPEG algorithm? The reward is the damaging file dimension of the picture (in kB) when saved as a JPEG.
• Incompressibility: How onerous is the picture to compress utilizing the JPEG algorithm? The reward is the constructive file dimension of the picture (in kB) when saved as a JPEG.
• Aesthetic High quality: How aesthetically interesting is the picture to the human eye? The reward is the output of the LAION aesthetic predictor, which is a neural community educated on human preferences.
• Immediate-Picture Alignment: How nicely does the picture characterize what was requested for within the immediate? This one is a little more sophisticated: we feed the picture into LLaVA, ask it to explain the picture, after which compute the similarity between that description and the unique immediate utilizing BERTScore.

Since Secure Diffusion is a text-to-image mannequin, we additionally want to select a set of prompts to present it throughout finetuning. For the primary three duties, we use easy prompts of the shape “a(n) [animal]”. For prompt-image alignment, we use prompts of the shape “a(n) [animal] [activity]”, the place the actions are “washing dishes”, “taking part in chess”, and “driving a motorcycle”. We discovered that Secure Diffusion typically struggled to provide photos that matched the immediate for these uncommon situations, leaving loads of room for enchancment with RL finetuning.

First, we illustrate the efficiency of DDPO on the straightforward rewards (compressibility, incompressibility, and aesthetic high quality). All the photos are generated with the identical random seed. Within the high left quadrant, we illustrate what “vanilla” Secure Diffusion generates for 9 totally different animals; the entire RL-finetuned fashions present a transparent qualitative distinction. Curiously, the aesthetic high quality mannequin (high proper) tends in the direction of minimalist black-and-white line drawings, revealing the sorts of photos that the LAION aesthetic predictor considers “extra aesthetic”.

Subsequent, we reveal DDPO on the extra complicated prompt-image alignment process. Right here, we present a number of snapshots from the coaching course of: every collection of three photos reveals samples for a similar immediate and random seed over time, with the primary pattern coming from vanilla Secure Diffusion. Curiously, the mannequin shifts in the direction of a extra cartoon-like model, which was not intentional. We hypothesize that it is because animals doing human-like actions usually tend to seem in a cartoon-like model within the pretraining information, so the mannequin shifts in the direction of this model to extra simply align with the immediate by leveraging what it already is aware of.

Surprising Generalization

Shocking generalization has been discovered to come up when finetuning giant language fashions with RL: for instance, fashions finetuned on instruction-following solely in English typically enhance in different languages. We discover that the identical phenomenon happens with text-to-image diffusion fashions. For instance, our aesthetic high quality mannequin was finetuned utilizing prompts that had been chosen from an inventory of 45 frequent animals. We discover that it generalizes not solely to unseen animals but in addition to on a regular basis objects.

Our prompt-image alignment mannequin used the identical record of 45 frequent animals throughout coaching, and solely three actions. We discover that it generalizes not solely to unseen animals but in addition to unseen actions, and even novel combos of the 2.

Overoptimization

It’s well-known that finetuning on a reward perform, particularly a discovered one, can result in reward overoptimization the place the mannequin exploits the reward perform to attain a excessive reward in a non-useful method. Our setting is not any exception: in all of the duties, the mannequin finally destroys any significant picture content material to maximise reward.

We additionally found that LLaVA is prone to typographic assaults: when optimizing for alignment with respect to prompts of the shape “[n] animals”, DDPO was capable of efficiently idiot LLaVA by as a substitute producing textual content loosely resembling the right quantity.

There’s presently no general-purpose technique for stopping overoptimization, and we spotlight this drawback as an essential space for future work.

Conclusion

Diffusion fashions are onerous to beat in relation to producing complicated, high-dimensional outputs. Nonetheless, to this point they’ve largely been profitable in functions the place the aim is to be taught patterns from heaps and plenty of information (for instance, image-caption pairs). What we’ve discovered is a solution to successfully prepare diffusion fashions in a method that goes past pattern-matching — and with out essentially requiring any coaching information. The chances are restricted solely by the standard and creativity of your reward perform.

The best way we used DDPO on this work is impressed by the current successes of language mannequin finetuning. OpenAI’s GPT fashions, like Secure Diffusion, are first educated on enormous quantities of Web information; they’re then finetuned with RL to provide helpful instruments like ChatGPT. Usually, their reward perform is discovered from human preferences, however others have extra lately found out produce highly effective chatbots utilizing reward features based mostly on AI suggestions as a substitute. In comparison with the chatbot regime, our experiments are small-scale and restricted in scope. However contemplating the large success of this “pretrain + finetune” paradigm in language modeling, it definitely looks like it’s value pursuing additional on the planet of diffusion fashions. We hope that others can construct on our work to enhance giant diffusion fashions, not only for text-to-image technology, however for a lot of thrilling functions resembling video technology, music technology,  picture modifying, protein synthesis, robotics, and extra.

Moreover, the “pretrain + finetune” paradigm just isn’t the one method to make use of DDPO. So long as you have got a great reward perform, there’s nothing stopping you from coaching with RL from the beginning. Whereas this setting is as-yet unexplored, it is a place the place the strengths of DDPO may actually shine. Pure RL has lengthy been utilized to all kinds of domains starting from taking part in video games to robotic manipulation to nuclear fusion to chip design. Including the highly effective expressivity of diffusion fashions to the combination has the potential to take current functions of RL to the following degree — and even to find new ones.

This submit is predicated on the next paper:

If you wish to be taught extra about DDPO, you may try the paper, web site, unique code, or get the mannequin weights on Hugging Face. If you wish to use DDPO in your personal mission, try my PyTorch + LoRA implementation the place you may finetune Secure Diffusion with lower than 10GB of GPU reminiscence!

If DDPO evokes your work, please cite it with:

@misc{black2023ddpo,
      title={Coaching Diffusion Fashions with Reinforcement Studying}, 
      creator={Kevin Black and Michael Janner and Yilun Du and Ilya Kostrikov and Sergey Levine},
      12 months={2023},
      eprint={2305.13301},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}