Release 0.35.0

Release Date: September 2026

Previous Version: v0.34.0

Build a character from a Krea 2 concept, bring it into 3D with Meshy 7.1, and render complete Blender scenes at 1080p. This release brings together the new features and improvements since 0.34.0, including the intervening patch releases. Model choices depend on your account's access and the node you use.

Getting started, canvas, and assistant

• Start directly on a canvas. Eligible new workspace owners start with a project and a prompt composer. Describe an asset or pipeline, attach references, and choose Build to hand it to the assistant. Template cards provide starting projects without running them automatically, and the starter project takes its name from your prompt. A dismissible mouse/trackpad hint replaces the blocking input-device prompt. This experience is rolling out to qualifying new accounts in their first empty workspace, on writable desktop canvases.

• Find models in a consistent, searchable picker. Generation nodes and the assistant group models by provider or family, with search, keyboard navigation, and clearer labels.

• Choose the project assistant's model. Selecting a supported vision- and tool-capable model no longer needs a separate model-selection permission. Choosing a model other than Gemini 3.8 Flash opens a warning about stability and unpredictable costs. Don't show again saves your preference to your account.

• Ask the assistant to organize part of your graph. It can inspect frame membership, create nested groups, and arrange selected nodes in dependency order while leaving unrelated canvas content in place. It can also edit frames without moving their nodes. These are project-assistant capabilities, also usable through existing MCP assistant conversations.

• Read field help by hovering over its name. Field-name tooltips also work on read-only field displays.

• See historical average node costs. The Node Pricing page adds Avg/run, calculated from successful, uncached runs over the previous 30 days and refreshed weekly when enough observations are available. Reference price ranges and the inputs that affect cost help explain the differences. These averages describe past usage; they are not a quote for your next run.

• Manage referrals from one sidebar card. Copy your referral link and see reward amounts, signups, paid conversions, and earned credits together. This replaces the separate Referral section in Billing.

Image, language, music, and video models

• Generate images with Krea 2. Text + Image → Image and its image-list variant add Krea 2 Large, Medium, Medium Turbo, and Turbo, with style references and model-specific creativity and style-strength controls. Large, Medium, and Medium Turbo accept up to ten style references; Turbo accepts up to three. References guide the style rather than guaranteeing the same subject identity.

• Combine Krea LoRAs and use more functional image controls. The new Krea 2 LoRA Selector combines up to three preset or custom LoRAs, each with its own strength and preview. Presets include ink wash and a contrast slider. Connect the selector to Krea 2 Turbo for text-based generation without reference images. The functional LoRA catalog also adds Base-to-Turbo and RefControl Depth options, with updated guidance and defaults. Requires LoRA access.

• Use Grok 4.7 and Claude Opus 5.5. Both join Atlas's language and vision model catalog for compatible nodes and assistant workflows. Opus 5.5 replaces the previous Opus choice, and saved older Opus selections migrate automatically.

• Create full songs with Lyria 3.5. Lyria 3.5 replaces Lyria 3 Pro, with saved Pro selections migrated to the new choice. Music generation gains two individual image-reference inputs alongside the existing image-list input, plus expanded composition-plan and lyrics guidance. Lyria 3 Clip remains available.

• Edit videos with FLUX 3. Video Edit adds prompt-guided changes to an existing video while retaining its motion, timing, and framing. The Atlas input limits are MP4 videos shorter than 15 seconds and smaller than 50 MB.

• Use Seedance 2.5 at native 1080p. Generation, reference generation, editing, and extension support native 1080p. Connecting a video reference exposes a Reference motion / style mode. Native runs settle against reported token usage. Atlas retains the higher-quality source video and provides a separate browser-compatible preview when needed.

3D generation, rigging, and animation

• Choose Meshy 7.1 or Meshy 6 Lite. Image-to-3D and multiview workflows gain both choices. Single-image Meshy 7.1 offers Standard, Ultra 2K, and Ultra 4K geometry detail, separate from texture resolution and face count, with triangle and quad variants. Meshy 6 Lite replaces the older v5 retexturing choice. Saved Meshy selections migrate where applicable.

• Split generated models with Meshy Auto Split. Split a Meshy model generated in the same project automatically or by naming its parts. Choose assembled or print-bed layout and optionally add cube or cylinder connectors with size controls. Generation nodes expose a Meshy Task ID for connecting this workflow. Auto Split does not support Ultra 4K geometry.

• Include multiple Meshy animations in one model. Animate Rigged Mesh accepts up to ten distinct catalog actions and returns a GLB with one animation clip per selection. The animation catalog also expands.

• Control Pixal3D geometry and textures. The existing Pixal3D image-to-3D option gains geometry resolution, texture resolution, and target-vertex controls, with remeshing and PBR textures.

• Generate character motion from text with Kimodo. Kimodo Create Character Animation creates 0.5–10-second clips on a Mixamo skeleton for character-animation workflows. Finger bones stay in a fixed, relaxed pose. Requires animation access.

Blender workflows

• Choose a render-quality preset. Render Blender Scene adds Fast, 1080p, and Custom / Saved Scene modes. Fast is the new default; Custom uses the saved scene settings and any enabled overrides. Default runtime and credit budgets increase. Camera auto-fit is off by default, preserving the scene's framing unless you enable it.

• Continue longer visual Blender conversations. Blender Agent uses a rolling window of recent images for each model completion, replacing the cumulative screenshot limit. Older images remain in the saved conversation. The default window is six images, though the platform configuration can change it.

• Catch oversized Blender inputs before spending credits. Blender Agent, Render Blender Scene, and Blender Scene to GLB check the combined input size before allocating a worker or spending credits. The default limit is 500 MB.