← CHI 2026 map

CHI '26 · Honorable mention · full-paper review · confidence medium-high

Noise Pilot: Enabling Artistic Workflow Composition with Diffusion-Based Image Generation

James Smith , Shm Garanganao Almeda , Timothy J. Aveni , Anya Agarwal , Bjoern Hartmann

DOI PDF Program page Page JSON

This is a strong CHI creativity-support paper because the contribution is not just another prompt interface: it exposes diffusion as editable workflow material and shows artists using that material in practice. The study is small, but the design idea is clear and well matched to the deployment evidence.

Axes Lens

Rare contribution shape, typical evidence profile. The point here is not a score. It is to show what kind of claim the paper makes, and whether the evidence pattern is unusual or baseline in this 268 -review set.

Review Summary

Noise Pilot reads as a thoughtful CHI contribution because it targets a real mismatch in current creative AI tools: artists are asked to steer diffusion systems through prompts while the actual generative process stays hidden. The paper’s core move is to make that process visible and editable, turning diffusion steps into composable nodes rather than an opaque backend. That is a meaningful shift in interaction design, not just a cosmetic interface change. The evidence also fits the claim structure reasonably well: the authors do not rely on a lab benchmark, but instead deploy the tool with 9 expert creative practitioners over two weeks and observe how they move across three interaction depths, build reusable artifacts, and sometimes achieve results that prompting alone could not produce. That makes the paper strongest as a design-and-practice argument about what becomes possible when diffusion is treated as a manipulable material. The main caution is scope. The study is qualitative and small, so it supports insight into expert creative workflows more than broad generalization. The system also has explicit constraints, including 256x256 output and no training-level control, which means it does not cover the full space of artist workflows. Even so, the paper’s contribution is coherent and well grounded: it offers a concrete alternative to black-box GenAI interfaces and demonstrates that artists can use that alternative in meaningful ways.

What Changed

Canon before

Prior CST and GenAI image tools largely expose diffusion-based generation as prompt-driven black boxes, with control either indirect through prompting or requiring engineering-heavy customization.

Departure from common sense

The paper argues against the common CST assumption that diffusion-based image generation should remain a black box controlled mainly by prompts. Instead, it treats the diffusion process itself as something artists should see and manipulate directly, so expressive control can come from composable workflow material rather than only from prompting.

Actual novelty

Noise Pilot’s novelty is to implement the standard diffusion process as editable nodes inside the tool, so users can customize the denoise loop and compose it with other image operations into reusable workflows. That shifts diffusion from an opaque backend into a directly manipulable artistic material.

Evidence

The paper presents a tool that makes diffusion steps visible and editable, then studies it in a 2-week deployment with 9 expert creative practitioners. Evidence supports both the design claim and the observed use of three interaction depths, including reusable artifacts and outcomes beyond prompting alone.

Limits

Method limits

The evidence comes from a small qualitative deployment with expert participants, so it supports design insight and observed practice more than broad causal claims or population-level generalization.

Deployment limits

Noise Pilot supported a maximum image resolution of 256x256 pixels and did not include training-level model manipulation such as dataset curation or LoRA-style adaptation.

Boundary conditions

The approach is most applicable to expert or highly motivated creative practitioners working with diffusion-based image generation, especially where workflow composition and process visibility matter more than turnkey simplicity.

Manual check: No additional manual check required for the extracted claims, but the small deployment scope should be kept in mind when interpreting generality.

Position in field

This positions the paper as a CHI creativity-support contribution that reframes GenAI image tools from prompt-centric black boxes toward visible, composable diffusion workflows, emphasizing material control over hidden model behavior.

Abstract

Creativity support tools (CSTs) increasingly include image-generation features. The underlying diffusion models enact a particular image diffusing process that AI CSTs tend to obscure within a black-box. Artists’ creative control is limited to indirect manipulation (prompting), chaining these "black-boxes" together, or using ML-engineering skills to build custom black-boxes. Seeking to maintain the low-threshold offered by prompting, while raising the ceiling of expressive interactions, we built Noise Pilot: a multi-layered approach to supporting diffusion-based creative processes at three levels of depth. We used Noise Pilot as a probe to study the artistic processes of 9 artists over a 2-week period. Artists engaged with diffusion at different levels of manipulative depth and crafted reusable artifacts to enact bespoke diffusion processes; some produced results impossible to achieve with prompting alone. We discuss how black-box AIs in CSTs limit creative power, and propose subverting this by favoring visibility over obscurity, and materiality over personification.