Open Problems

Diffusion Models

Black-Box and Few-Step Control of Diffusion Models Without Intermediate State Access

Barrier to removeOpen
Strong candidate · 4/5 runs6 papers report this50% from 2025+

Generated automatically from the limitations stated in 6 papers (ICLR, ICML, NeurIPS), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.

The problem

Existing diffusion manipulation, guidance, and acceleration methods almost universally require white-box model access, intermediate noise predictions, or dozens of trajectory steps. Consequently, these techniques fail when interacting with commercial API endpoints that only return final image outputs, or with modern few-step distilled models (e.g., 1-to-4-step generators) where temporal redundancy across timesteps does not exist. Practitioners cannot apply state-of-the-art guidance, watermarking, or editing pipelines to closed-source foundation models or ultra-fast distilled backends. Resolving how to steer or audit generation under sample-only access is required to bridge the gap between white-box research methods and deployed diffusion infrastructure.

Why it matters

Enables downstream steering, safety filtering, and controllable generation on proprietary commercial APIs and ultra-fast distilled diffusion models that do not expose intermediate representations.

Ways to approach it

Prior-work checks are free with an account. Results someone already ran are shown to everyone.

  1. 1

    Construct a standardized benchmark evaluating existing editing and inversion techniques across decreasing levels of internal access: from full white-box latents down to black-box API queries (seed/prompt to image) and 1–4 step distilled models (e.g., Flux-schnell, SDXL-Turbo), measuring fidelity drop (CLIP score, LPIPS) and query budget.

  2. 2

    Develop a zeroth-order sample-space feedback mechanism that optimizes conditioning prompts or input noise distributions using only final sample evaluations, measuring convergence rate and perceptual quality against white-box score-based guidance baselines.

  3. 3

    Formulate post-hoc transformation pipelines that operate directly on generated sample pairs to achieve target attribute modifications without relying on intermediate reverse-ODE trajectories, measuring identity preservation and edit success rate.

Have a different approach?

Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.

Free · 3 checks per day

Why it might fail

If API providers begin standardizing open access to intermediate latents and token embeddings, or if black-box sample-level query optimization proves too sample-inefficient and costly compared to running local open-weights models.

Evidence

Each paper's own statement of the limitation, verbatim.

Nearest existing work

Related open problems

Diffusion Models

Scope to testOpen

Extending Class-Conditional ImageNet Generation Methods to Text-Conditioned, Open-Domain Synthesis

Nearly every method validated on class-conditional ImageNet 256×256 stalls at that benchmark, and it is unknown whether any of them survive contact with natural language conditioning and open-caption distributions. Class labels provide weak, discrete conditioning with 1,000 modes; text requires cross-attention or caption-token conditioning, resolution diversity, and open-vocabulary semantics. Until the class-to-text transition is demonstrated, the large literature of ImageNet-benchmarked generative advances cannot be assumed to transfer to practical text-to-image systems.

Possible candidate · 3/5 runs53 papers report this83% from 2025+
Generated automatically, not curated by hand. Automated prior-work checks catch about a third of existing work, so treat this problem as a lead to investigate.