Methods worth understanding

AI Techniques Radar

New AI methods, the evidence behind them, and how to put them to work.

Latest weekly report

AI Techniques Weekly

Methods worth understanding, their results, and practical limitations.

Read the report
Reset

Latest discoveries · 115

Last refreshed 2026-11-10 · Some sources need attention

Training & fine-tuning

Fine-tune an LLM for RLHF by keeping the base model in 8-bit and training only low-rank adapter weights (via PEFT) while computing both active and reference model logits using the same model by temporarily disabling adapters (so two full copies are not required).

AI summary

Inference efficiency

Convert a Stable Diffusion pipeline to OpenVINO via Optimum Intel, then reshape the exported pipeline to a fixed batch size and fixed image height/width when your application only uses a small set of resolutions. This enables the runtime to use a static input shape rather than dynamic shapes.

AI summary

Training & fine-tuning

Adapt a text-to-image diffusion model for image-to-image editing by instruction-tuning so it can take an input image plus an explicit instruction phrase and produce the edited result that matches the transformation specified in the prompt (e.g., cartoonization or deraining).

AI summary

Training & fine-tuning

Train a Stable Diffusion pipeline with quantization-aware training (QAT) using NNCF, augment the training objective with knowledge distillation (teacher-student setup using the source model as teacher), stabilize training with EMA of parameters (excluding quantizers), and—during optimization—modify the attention UNet by applying token merging before self-attention and stack it with 8-bit quantization for further efficiency.

AI summary

Inference efficiency

Compress Stable Diffusion model weights using 6-bit weight palettization (a form of quantization) during conversion to the Core ML format, then keep weights compressed as palette indices and convert (de-palettize) them on the fly as inference progresses layer by layer rather than fully uncompressing at model load time.

AI summary

Training & fine-tuning

The approach fine-tunes a Stable Diffusion text-to-image model using textual inversion, where a new placeholder token is learned from a very small set of example images. In the described setup, the learnable token is trained against five images to associate the token with a specific visual concept, producing a saved model suitable for later image generation.

AI summary

Training & fine-tuning

Mixed-bit palettization is a post-training model compression approach that assigns different quantization bit depths per layer (e.g., 1/2/4/8 bits, with some layers kept at higher precision) based on each layer’s estimated contribution to quality degradation, producing a JSON “recipe” that is later applied to replace layer weights with palettized versions.

AI summary

Training & fine-tuning

Use a two-stage image compression/decoder (Stage A + Stage B) to map images into an extremely compressed latent space, then train a separate generative model (Stage C) to produce those latents conditioned on text. The decoder modules map the generated latents back to pixel space.

AI summary

Training & fine-tuning

Denoising Diffusion Policy Optimization (DDPO) reframes diffusion denoising as a multi-step Markov decision process and fine-tunes the diffusion model using a PPO-style policy-gradient update, where a reward is applied at the end of the denoising trajectory. In an RLHF-like workflow for image generation, a reward model (e.g., an aesthetic predictor) scores generated samples, and those rewards drive DDPO to adjust the diffusion model toward the objective captured by the reward.

AI summary

Inference efficiency

Compile an SDXL image-generation pipeline in JAX with just-in-time (JIT) compilation that assumes fixed (static) input/output shapes, then scale generation across multiple TPU chips using pmap-style parallelism so multiple images render concurrently with the same compiled program.

AI summary

Inference efficiency

Use a faster “assistant” language model to propose a block of candidate tokens, then verify them with a slower “main” model using a single forward pass over the block. Keep tokens up to the first mismatch with the main model, replace the mismatching token with the main model’s choice, and discard the remaining candidates; repeat until decoding completes. The method is described as producing exactly the same outputs as running the main model alone.

AI summary