| name | consistency-solver-preview-refine |
| title | Image Diffusion Preview with Consistency Solver: Trainable Adaptive ODE Solver for Interactive Generation |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2512.13592 |
| keywords | ["diffusion","interactive-generation","preview-refine","ODE-solver","reinforcement-learning"] |
| description | Train adaptive ODE solvers that generate fast diffusion previews while maintaining consistency with full-step refinement. Learn context-aware integration coefficients through PPO without distilling base models. Achieve high-quality previews in few steps with 47% fewer steps than standard methods. |
Skill Summary
ConsistencySolver is a trainable, adaptive ODE solver designed for efficient image generation through a preview-and-refine workflow. Rather than distilling models (which alters weights), it learns context-aware integration coefficients during diffusion sampling via reinforcement learning. The approach preserves original model properties while optimizing sampling trajectory, reducing overall interaction time by approximately 50% through fast low-step previews with consistent full-step refinement.
When To Use
- Interactive image generation applications requiring preview-then-refine workflows
- Scenarios where fast low-quality previews guide high-quality final generation
- Projects where preserving original diffusion model properties is important
- User-facing applications where interaction latency dominates total time
When NOT To Use
- Single-pass generation scenarios where preview-refine overhead isn't justified
- Applications already using distilled fast models with acceptable quality
- Contexts where modifying sampling trajectory causes artifacts
- Scenarios with tight token/compute budgets preventing RL training
Core Technique
ConsistencySolver combines three key ideas:
1. Learnable Multistep ODE Solver
Rather than fixed numerical schemes, learn context-aware weights through RL (PPO). The solver formula is:
y_{t_{i+1}} = y_{t_i} + (n_{t_{i+1}} - n_{t_i}) · [∑j w_j(t_i, t{i+1}) · ε_{i+1-j}]
where adaptive coefficients are generated by a lightweight neural network based on timestep information.
2. Preview-and-Refine Paradigm
Enable fast, low-step preview generation that remains consistent with full-step refinement outputs. Unlike distillation, preserve the original diffusion model's properties while optimizing the sampling trajectory.
3. Reinforcement Learning Training
Use RL training instead of costly distillation, achieving "FID scores on-par with Multistep DPM-Solver using 47% fewer steps." Learn adaptation without weight modification.
Implementation Notes
Implement lightweight neural network that generates timestep-dependent integration coefficients. Train via PPO to maximize preview quality consistency with full refinement. Integrate into diffusion sampling loop to produce fast previews. Use human evaluation to optimize preview-refine balance for your application.
References
- Original paper: Image Diffusion Preview with Consistency Solver (Dec 2025)
- ODE solvers for diffusion models
- Reinforcement learning for sampling optimization
