MAJOR Upgrades to AI Art Tools! & Open AI's new Image Gen!

TL;DR

Idiogram Canvas adds a fully interactive editing workflow: drag variations, zoom, delete unwanted results, and continue refining on a shared canvas.

Briefing Cornell Notes

Briefing

AI image editing is shifting from “generate and pick” toward real, Photoshop-like workflows—led by Idiogram Canvas. The standout upgrade is a fully interactive canvas that lets users drag, zoom, delete unwanted variations, and then extend images seamlessly using an outpainting-style workflow. In a hands-on example, a Halloween-costumed cat is generated with multiple variations, then the canvas is expanded to a wider 3:2 aspect ratio. With “Magic Prompt” turned off, the extend tool produces a background that matches the prompt intent (a suburban Halloween street) without inventing extra details. A separate “Magic Fill” tool adds localized edits: using lasso masking, a speech bubble is generated with text like “dang these people are giving away candy not cat treats,” and the bubble appears in the correct region.

The practical value is that these tools behave like an editing suite rather than a one-shot generator. Users can keep the original image while extending, choose among four results, and continue refining with additional fills. The workflow also supports image uploading and remixing—so a user can extend a personal photo into a new scene (e.g., lemon-themed headphones, lemon-themed styling) or expand a small Halloween cat into a full pumpkin-filled background. The main limitations noted are prompt sensitivity (vague prompts like “background” can lead to unintended changes, including adding other people) and occasional lag/lower frame rate. Still, automatic cloud saving and the overall UI responsiveness make it feel close to a daily-driver editor for many tasks.

Midjourney is also moving toward deeper editing. New tests include an image editor for uploaded images that performs inpainting (erase and regenerate a selected area) and “image retexturing,” which changes materials, surfacing, and lighting while preserving the underlying composition. Examples show a dog with a party hat added after erasing part of the image, a body erased and replaced with a colorful sweater that still fits the same animal, and interior photo retexturing that transforms a living room into a far-future version while keeping key elements like couches and books in place. The canvas extension exists too, but it’s described as more constrained than Idiogram’s free-form approach—potentially easier for new users.

Beyond editing tools, the model landscape keeps shifting. Stability AI released Stable Diffusion 3.5 Large, positioning it as a more usable update to Stable Diffusion 3 after earlier reception. It’s framed as more open and easier to fine-tune, with community discussion pointing to built-in support for training and control mechanisms like LoRA-style workflows, IP-Adapter use, and fine-tuning. OpenAI introduced “consistency models” (SCMS), designed to generate samples in only two sampling steps—faster than diffusion—while reporting quality comparable to leading diffusion approaches. The roundup closes with the idea that OpenAI is still actively pushing image generation research, even after major releases like DALL·E 3.

Taken together, the message is clear: the competitive edge is moving from raw image quality alone toward controllable, iterative editing—where extending, masking, and retexturing can be chained into a coherent creative workflow.

Cornell Notes

Idiogram Canvas is presented as a major leap in AI image editing, combining drag-and-drop canvas control with extend (outpainting) and magic fill (masked inpainting). The extend tool can expand an image to new aspect ratios and generate seamless backgrounds, especially when “Magic Prompt” is disabled to prevent unwanted inference. Magic Fill uses brush or lasso masks to add localized elements like speech bubbles, and the workflow supports uploaded images for iterative edits. Midjourney’s parallel updates—uploaded-image inpainting and “image retexturing”—aim to preserve composition while changing materials, lighting, and even style. The broader model news includes Stable Diffusion 3.5 Large and OpenAI’s SCMS consistency models, both pushing speed, fine-tuning flexibility, and practical usability.

What makes Idiogram Canvas feel more like an editor than a generator?

It adds a true interactive canvas: users can drag and reposition generated variations, delete unwanted ones without losing the rest, zoom in via control+scroll, and then extend the image to a new aspect ratio. The extend tool outpaints the scene while keeping the original image available, and Magic Fill applies localized edits using masks (brush, lasso, or rectangle). The workflow is iterative—extend, then fill, then refine—rather than a single generation-and-pick loop.

Why does turning “Magic Prompt” off matter when extending an image?

With “Magic Prompt” off, the extend tool follows the prompt more literally and avoids adding extra inferred details. The transcript highlights that leaving it on can lead to unwanted additions, while disabling it helps produce a background that matches a straightforward prompt like “a typical Suburban Street on a Halloween night.”

How does Magic Fill differ from the extend tool in practical use?

Extend changes the canvas by expanding the scene (outpainting), while Magic Fill edits a specific region inside the existing image. Magic Fill relies on masking—like a lasso selection—to target where new content should appear. In the example, a lasso mask selects an area and a speech bubble is generated with text, producing a localized overlay rather than a full-scene expansion.

What new Midjourney capabilities are highlighted, and what do they preserve?

Midjourney’s tests include inpainting for uploaded images (erase part of an image and regenerate it) and “image retexturing,” which changes materials/surfacing/lighting while keeping the original composition. Examples include erasing and replacing a dog’s body with a sweater that still fits the same dog, and retexturing a room into a far-future scene while retaining key elements like couches and books.

What’s the significance of Stable Diffusion 3.5 Large and SCMS in the broader model race?

Stable Diffusion 3.5 Large is framed as a redemption of Stable Diffusion 3, emphasizing improved usability and openness, plus easier fine-tuning. OpenAI’s SCMS (consistency models) targets speed by generating samples in only two sampling steps, aiming for quality comparable to leading diffusion models while reducing generation time.

Review Questions

When would a user prefer “Magic Prompt” off in Idiogram Canvas, and what failure mode does it help avoid?
Compare Idiogram’s extend+Magic Fill workflow with Midjourney’s inpainting+retexturing: which tasks map to each toolset?
What tradeoffs are implied between faster sampling (SCMS) and diffusion quality, based on the examples described?

Key Points

1
Idiogram Canvas adds a fully interactive editing workflow: drag variations, zoom, delete unwanted results, and continue refining on a shared canvas.
2
The extend tool can outpaint to new aspect ratios, and disabling “Magic Prompt” helps prevent the system from inventing extra details beyond the prompt.
3
Magic Fill enables localized edits through masking (brush/lasso/rectangle), making it suitable for adding specific elements like speech bubbles.
4
Midjourney’s new uploaded-image editor focuses on inpainting and “image retexturing,” preserving composition while changing materials, lighting, and style.
5
Stable Diffusion 3.5 Large is positioned as a more usable update to Stable Diffusion 3 with improved fine-tuning flexibility and openness.
6
OpenAI’s SCMS consistency models aim to cut generation time by using only two sampling steps while targeting diffusion-level quality.
7
Across tools, prompt specificity remains crucial: vague prompts can cause unintended edits even when the UI feels seamless.

Highlights

Idiogram Canvas’s extend workflow can expand a generated image to a wider aspect ratio and produce a seamless background—especially when “Magic Prompt” is turned off.

Magic Fill’s lasso masking can generate a speech bubble in the selected region, demonstrating localized editing rather than full-scene regeneration.

Midjourney’s “image retexturing” can transform a real interior photo into a far-future environment while keeping the original furniture layout intact.

OpenAI’s SCMS consistency models target speed by generating samples in only two sampling steps, trading off some quality for faster results.

Topics

Idiogram Canvas
Midjourney Inpainting
Image Retexturing
Stable Diffusion 3.5 Large
OpenAI SCMS