RectifiedHR: Ankit Sanjyal

RectifiedHR: High-Resolution Diffusion via Energy Profiling and Adaptive Guidance Scheduling

Ankit Sanjyal

arXiv preprint arXiv:2507.09441 · cs.GR, cs.CV · July 2025 (v2: December 2025)

Diffusion Models High-Resolution Synthesis Classifier-Free Guidance Computer Vision

arXiv PDF Code

RectifiedHR teaser: adaptive CFG scheduling results

Abstract. High-resolution image synthesis with diffusion models often suffers from energy instabilities and guidance artifacts that degrade visual quality. We analyze the latent energy landscape during sampling and propose adaptive classifier-free guidance (CFG) schedules that maintain stable energy trajectories. Our approach introduces energy-aware scheduling strategies that modulate guidance strength over time, achieving superior stability scores (0.9998) and consistency metrics (0.9873) compared to fixed-guidance approaches. We demonstrate that DPM++ 2M with linear-decreasing CFG scheduling yields optimal performance, providing sharper, more faithful images while reducing artifacts. Our energy profiling framework serves as a powerful diagnostic tool for understanding and improving diffusion model behavior.

Motivation

Standard diffusion models use a fixed classifier-free guidance scale (CFG) throughout all denoising timesteps. This is simple but suboptimal: early timesteps (coarse structure) and late timesteps (fine details) have very different signal-to-noise regimes, yet both receive the same guidance push. The result is guidance artifacts: oversaturation, halos, and energy spikes: that become more visible at higher resolutions.

RectifiedHR asks: What does the latent energy landscape actually look like during sampling, and can we schedule guidance to keep it stable?

Method

Energy Profiling

We profile the L2 norm of the latent tensor at each denoising step for a fixed set of prompts across multiple CFG schedules (constant, linear-increasing, linear-decreasing, cosine, step). This gives us an empirical energy trajectory: a diagnostic fingerprint of each schedule's stability.

Adaptive CFG Schedules

Rather than a fixed CFG scale w, we define a time-varying schedule w(t). The key insight is that linear-decreasing guidance: starting high to lock in structure and tapering off as details solidify: best matches the natural energy dynamics of the denoising process. We evaluate four schedule families:

Constant: baseline, fixed w throughout
Linear-decreasing: w anneals from w_max to w_min (best performer)
Linear-increasing: w ramps up over timesteps
Cosine: smooth sinusoidal modulation

Sampler

We pair adaptive CFG with DPM++ 2M, a second-order multi-step solver that is particularly sensitive to guidance quality. This combination yields the most stable energy trajectories and the best perceptual results.

Results

Schedule	Stability Score ↑	Consistency ↑	Artifacts
Constant CFG (baseline)	0.9821	0.9614	Visible at 512+
Linear-increasing	0.9877	0.9702	Moderate
Linear-decreasing (DPM++ 2M)	0.9998	0.9873	Minimal
Cosine	0.9934	0.9801	Low

Stability score = 1 − normalized variance of latent energy across steps. Higher is better.

Key Takeaways

Fixed CFG is a suboptimal heuristic: the energy landscape is highly non-uniform across timesteps.
Linear-decreasing schedules naturally align guidance with the coarse→fine denoising progression.
The energy profiling framework is model-agnostic and can diagnose any diffusion pipeline.
No retraining required: this is a pure inference-time improvement.

Citation

@article{sanjyal2025rectifiedhr,
  title     = {RectifiedHR: High-Resolution Diffusion via Energy Profiling and Adaptive Guidance Scheduling},
  author    = {Sanjyal, Ankit},
  journal   = {arXiv preprint arXiv:2507.09441},
  year      = {2025}
}