Coaching Diffusion Fashions with Reinforcement Studying – The Berkeley Synthetic Intelligence Analysis Weblog

Coaching Diffusion Fashions with Reinforcement Studying

Diffusion fashions have just lately emerged because the de facto customary for producing complicated, high-dimensional outputs. Chances are you’ll know them for his or her means to provide stunning AI art and hyper-realistic synthetic images, however they’ve additionally discovered success in different functions akin to drug design and continuous control. The important thing concept behind diffusion fashions is to iteratively rework random noise right into a pattern, akin to a picture or protein construction. That is usually motivated as a maximum likelihood estimation drawback, the place the mannequin is skilled to generate samples that match the coaching knowledge as carefully as attainable.

Nevertheless, most use instances of diffusion fashions are usually not instantly involved with matching the coaching knowledge, however as a substitute with a downstream goal. We don’t simply need a picture that appears like current photographs, however one which has a selected sort of look; we don’t simply need a drug molecule that’s bodily believable, however one that’s as efficient as attainable. On this publish, we present how diffusion fashions could be skilled on these downstream aims instantly utilizing reinforcement studying (RL). To do that, we finetune Stable Diffusion on a wide range of aims, together with picture compressibility, human-perceived aesthetic high quality, and prompt-image alignment. The final of those aims makes use of suggestions from a large vision-language model to enhance the mannequin’s efficiency on uncommon prompts, demonstrating how powerful AI models can be used to improve each other with none people within the loop.

diagram illustrating the RLAIF objective that uses the LLaVA VLM

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

Denoising Diffusion Coverage Optimization

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

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

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

chart comparing DDPO with RWR

We consider two variants of DDPO and two variants of RWR on three reward capabilities and discover that DDPO persistently achieves the perfect 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 the whole sequence of denoising steps that received us there. To do that, we reframe the diffusion course of as a multi-step Markov decision process (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 an alternative of utilizing the approximate chance of the ultimate pattern, these algorithms use the precise chance of every denoising step, which is extraordinarily simple to compute.

We selected to use coverage gradient algorithms because of their ease of implementation and past success in language model finetuning. This led to 2 variants of DDPO: DDPOSF, which makes use of the straightforward rating perform estimator of the coverage gradient also referred to as REINFORCE; and DDPOIS, which makes use of a extra highly effective significance sampled estimator. DDPOIS is our best-performing algorithm and its implementation carefully follows that of proximal policy optimization (PPO).

Finetuning Secure Diffusion Utilizing DDPO

For our predominant outcomes, we finetune Stable Diffusion v1-4 utilizing DDPOIS. We have now 4 duties, every outlined by a unique reward perform:

  • Compressibility: How simple is the picture to compress utilizing the JPEG algorithm? The reward is the damaging file measurement of the picture (in kB) when saved as a JPEG.
  • Incompressibility: How exhausting is the picture to compress utilizing the JPEG algorithm? The reward is the optimistic file measurement 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 skilled on human preferences.
  • Immediate-Picture Alignment: How effectively does the picture signify 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 choose a set of prompts to offer 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”, “enjoying chess”, and “using a motorcycle”. We discovered that Secure Diffusion usually struggled to provide photographs 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). The entire photographs are generated with the identical random seed. Within the prime 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 (prime proper) tends in direction of minimalist black-and-white line drawings, revealing the sorts of photographs that the LAION aesthetic predictor considers “extra aesthetic”.

results on aesthetic, compressibility, and incompressibility

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 sequence of three photographs exhibits samples for a similar immediate and random seed over time, with the primary pattern coming from vanilla Secure Diffusion. Curiously, the mannequin shifts in direction of a extra cartoon-like fashion, which was not intentional. We hypothesize that it’s because animals doing human-like actions usually tend to seem in a cartoon-like fashion within the pretraining knowledge, so the mannequin shifts in direction of this fashion to extra simply align with the immediate by leveraging what it already is aware of.

results on prompt-image alignment

Surprising Generalization

Stunning generalization has been discovered to come up when finetuning giant language fashions with RL: for instance, fashions finetuned on instruction-following solely in English often improve in other 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 have been chosen from a listing of 45 widespread animals. We discover that it generalizes not solely to unseen animals but additionally to on a regular basis objects.

aesthetic quality generalization

Our prompt-image alignment mannequin used the identical listing of 45 widespread animals throughout coaching, and solely three actions. We discover that it generalizes not solely to unseen animals but additionally to unseen actions, and even novel mixtures of the 2.

prompt-image alignment generalization


It’s well-known that finetuning on a reward perform, particularly a realized 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 isn’t any exception: in all of the duties, the mannequin ultimately destroys any significant picture content material to maximise reward.

overoptimization of reward functions

We additionally found that LLaVA is vulnerable 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.

RL exploiting LLaVA on the counting task

There’s at the moment no general-purpose methodology for stopping overoptimization, and we spotlight this drawback as an essential space for future work.


Diffusion fashions are exhausting to beat with regards to producing complicated, high-dimensional outputs. Nevertheless, up to now they’ve largely been profitable in functions the place the purpose is to study patterns from heaps and many knowledge (for instance, image-caption pairs). What we’ve discovered is a method to successfully prepare diffusion fashions in a method that goes past pattern-matching — and with out essentially requiring any coaching knowledge. The probabilities are restricted solely by the standard and creativity of your reward perform.

The way in which 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 skilled on enormous quantities of Web knowledge; they’re then finetuned with RL to provide helpful instruments like ChatGPT. Sometimes, their reward perform is realized from human preferences, however others have extra recently found out the right way to produce highly effective chatbots utilizing reward capabilities primarily based on AI suggestions as a substitute. In comparison with the chatbot regime, our experiments are small-scale and restricted in scope. However contemplating the big success of this “pretrain + finetune” paradigm in language modeling, it definitely looks like it’s value pursuing additional on the earth of diffusion fashions. We hope that others can construct on our work to enhance giant diffusion fashions, not only for text-to-image era, however for a lot of thrilling functions akin to video generation, music generation,  image editing, protein synthesis, robotics, and extra.

Moreover, the “pretrain + finetune” paradigm shouldn’t be the one method to make use of DDPO. So long as you have got a very good reward perform, there’s nothing stopping you from coaching with RL from the beginning. Whereas this setting is as-yet unexplored, this can be a place the place the strengths of DDPO might actually shine. Pure RL has lengthy been utilized to all kinds of domains starting from playing 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 subsequent stage — and even to find new ones.

This publish relies on the next paper:

If you wish to study extra about DDPO, you possibly can take a look at the paper, website, original code, or get the model weights on Hugging Face. If you wish to use DDPO in your individual challenge, take a look at my PyTorch + LoRA implementation the place you possibly can finetune Secure Diffusion with lower than 10GB of GPU reminiscence!

If DDPO evokes your work, please cite it with:

      title={Coaching Diffusion Fashions with Reinforcement Studying}, 
      writer={Kevin Black and Michael Janner and Yilun Du and Ilya Kostrikov and Sergey Levine},

Leave a Reply

Your email address will not be published. Required fields are marked *