AI Image Editing Is a Mess? He Put the Whole Creative Process on an Infinite Canvas

When AI Image Tools Become Slot Machines

Generating images with AI often feels like playing a slot machine: type a sentence, get four images, pick one, type it again. Twenty minutes later, your downloads folder is stuffed with final_v3_really_final.png, and you have no idea which prompt produced which image.

This pain point gave rise to a product called DREAMMMER. Its approach isn't to build "a better prompt input box" but to put the chat on the left and an infinite canvas on the right, where every request, prompt, render result, and revision hangs as a card from the node it originated from.

The Core Problem: AI Image Generation Lacks "Blood Relations"

People making product shots or social thumbnails rarely get it right in one pass. First adjust the lighting, then change the crop, then decide to put the bottle on a rock instead of a table. Every adjustment is a brand-new generation, and most tools treat them as unrelated, independent tasks.

What the author wanted was a kind of lineage tracking: hover over an early render and branch directly off that image — edit it, remove the background, upscale the resolution, or attach the image to the agent to continue the conversation — all without losing the full path that got you there.

The canvas is designed exactly for this. Searches connect to prompts, prompts connect to renders, and revisions hang beneath the image they modified. After ten rounds of iteration, you can still see exactly how each image came to be.

Lineage tracking in computer and information management is typically called lineage tracking or provenance tracking, referring to a system recording the full-chain information of where each object came from and what transformations it went through. In data engineering, it's used to trace which original tables a table's fields were generated from through which SQL transformations; in machine learning, it records which data, hyperparameters, and code trained a given model version. For AI image generation, lineage tracking means every render carries a reference to its parent node, which might be a previous image, a prompt, or a turn of an agent's reply. The value lies in reproducibility and traceability: you can follow the chain to find the version before a certain change without relying on filenames or memory. In implementation, it's usually stored as a directed acyclic graph, where nodes are assets or operations and edges represent derivation relationships — which is also the data structure commonly adopted by canvas-style products.

A Typical Workflow

The author describes their actual path of operation:

1. Describe the idea in plain language — no need to write a Midjourney novel
2. The agent converts it into a detailed English prompt and automatically selects a model (Auto mode), or you can manually lock in a specific model
3. Continue iterating in natural language: cooler lighting, simpler composition, keep the product label unchanged
4. Once satisfied with the static image, sometimes ask for a short video directly in the same conversation, rather than exporting a file and jumping to another app

The surprise of that last step even exceeds the canvas itself — going from image to subtle motion without rebuilding context elsewhere.

Auto mode in the article refers to the system automatically choosing a generation model based on the user's request, rather than the user specifying one manually. The image generation field has many models, each with different strengths in style and tasks: some lean toward realistic photographic quality, some excel at illustration and anime, and some are more stable for text layout or product consistency. A common implementation of Auto mode is to have a smaller language model first parse the user's intent and constraints, then route the request to the most suitable model based on preset rules or scoring, with corresponding parameter adjustments. Its advantage is reducing choice overhead; its disadvantage is that when a certain type of task requires a specific model to get the ideal result, automatic selection may produce a biased outcome — so mature products usually also keep a manual switch entry point. This aligns with the compromise mentioned in the article.

One-Click Tools Still Have a Place

Not every operation needs a conversation. Sometimes you just want to remove a background or restore an old photo. These features live on separate tool pages: upload, pick a tool, get the result. What's genuinely useful is that the result still opens in the chat window, where you can continue refining with the agent instead of starting over.

The Credit System

New accounts get 50 free credits. Consumption depends on the selected model and parameter settings (resolution, video length, etc.). If a generation fails, the credits consumed by that task are refunded. Specific rates are published on the pricing page, so users don't have to guess their balance mid-generation.

Unsolved Design Challenges

Interaction design for canvas-style products is hard to get right. Too many nodes and it looks like a plate of spaghetti; too aggressive an auto-layout and people feel they've lost control. The author is iterating on snap alignment, connection path routing, and the chat/canvas split — users can drag the divider or take the canvas area fullscreen.

Another issue is automatic model selection: Auto mode is good enough most of the time, but some tasks need a specific model to get that "right flavor." Offering both Auto and a manual model menu is the compromise found so far.

Auto-layout for canvas-style products refers to the system automatically calculating node positions to avoid overlap and tangled connections without breaking the user's mental model. Common approaches include force-directed layout, hierarchical layout, and incremental snapping: when a new node appears, it first tries a nearby free area, then nudges existing nodes slightly aside rather than rearranging the whole graph. The difficulty is that the sense of control and neatness are in natural conflict — the more aggressive the rearrangement, the cleaner the graph, but the positions the user just memorized get changed, their spatial memory fails, and it becomes harder to find a particular image again. The usual compromise is to do only local pushing or snap alignment, offer manual position locking, and perform global tidying only when the user actively triggers it. As a result, the interaction for such products often needs multiple rounds of iteration to reach a usable state.

Who Should Try It

If you've been ping-ponging between Midjourney, Flux, and various video tools, and you're tired of re-explaining the same creative brief every time, then the "conversation + history canvas" combination may be worth a look. The author is also soliciting feedback: what would make this way of working trustworthy enough that you'd use it end-to-end on a client project?

How to use it: log in with a Google account. Browsing templates and public portfolios doesn't require an account, but creating content does, so that conversation history and credits are bound to your personal account.

One-line Summary

DREAMMMER uses chat plus an infinite canvas to create lineage tracking for AI image generation, solving version and prompt traceability challenges.

Paragraph Summary

DREAMMMER addresses the slot-machine-like, hard-to-trace nature of AI image generation by putting chat on the left and an infinite canvas on the right. Every request, prompt, render, and revision hangs as a card from its source node, forming a traceable lineage: searches connect to prompts, prompts connect to renders, and revisions hang beneath the image they modified. Users can iterate in natural language, with the agent converting it to English prompts and selecting models automatically or manually, and can even generate short videos directly from static images within the same conversation. Results from one-click tools like background removal and photo restoration return to the chat window for further refinement. New accounts get 50 free credits, and failed tasks are refunded. The author is still iterating on canvas layout and automatic model selection, and is soliciting feedback.