> For the complete documentation index, see [llms.txt](https://docs.atlas.design/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.atlas.design/atlas-ai-studio-overview/release-notes/release-0.34.0.md). # Release 0.34.0 Release Date: September 2026 Previous Version: v0.33.0 #### Project assistant * **The project assistant can inspect images directly.** Ask it to describe image content, compare images, or read legible text in images available to the project or workspace. It can inspect up to four images at once, including when "Run nodes" is disabled. SVG and EXR assets use their rendered previews; animated images use their first frame. The assistant's model selector now offers vision-capable agent models, and incompatible saved selections fall back to an eligible model. Available model choices remain subject to account permissions.
* **See what the project assistant is waiting for.** Chat distinguishes an active agent run from waiting for subscribed nodes and shows live node-status chips, including completion markers. Click a chip to locate its node in the editable graph. The session status response also exposes `subscribed_nodes`, with each node's ID and finished state, for clients that consume agent status.
* **Clearer assistant credit usage.** Chat now displays **“Credits: X used / Y limit.”** The explanation clarifies that the amount covers the pool started by the latest manual prompt or MCP call, including automatic follow-ups. This replaces the ambiguous Current Run/Last Run presentation without changing the billing policy.
#### New Models for Image and 3D Tools * **Create and edit Blender scenes with Blender Agent.** Start from a prompt, an existing `.blend` scene, a saved conversation, or a combination. Supply up to **16 GLB meshes** and **six reference images**, and choose an eligible vision-capable agent model. The node returns an editable Blender scene and a saved conversation, with optional GLB export. Advanced credit and runtime limits bound the work; useful completed scene checkpoints can be returned if a limit interrupts the run. Scene state and conversation history are independent inputs, so continuing an edit can reuse both. *Requires standard node permissions and configured service access.*
* **Blender Agents default to GPT-6 Astra and accept more conversation history.** The default model for newly created Blender Agent nodes is **GPT-6 Astra**. The conversation text allowance doubles from **24 KB to 48 KB**, covering the prompt, history, generated tool code, and tool results. Existing explicit model selections are preserved.
* **Render Blender scenes directly.** The new **Render Blender Scene** node produces a PNG using Cycles without an AI editing step. By default it uses the scene's saved camera, render dimensions, sample count, and frame. Override width, height, samples, frame, or camera independently; optional Auto-fit Camera helps frame the scene. Uploaded/generated `.blend` files also gain rendered previews. *Requires standard node permissions and configured service access.*
* **Convert Blender scenes to GLB.** The new **Blender Scene to GLB** node turns a saved or uploaded `.blend` scene into a mesh for downstream 3D nodes or download, without asking an agent to edit it. GLB export does not bake procedural materials or preserve every Blender feature; retain the `.blend` file for the complete scene. *Requires standard node permissions and configured service access.*
* **Pass Blender scenes and saved conversations through workflow APIs.** New scene input and API input/output nodes support `.blend` files. API Input Conversation and API Output Conversation support saved `.atlas-context.json` files, allowing a workflow to accept prior agent history and return updated history. Scene and conversation API inputs can be optional without a default file, supporting both first-run and continuation requests.*Requires standard node permissions and configured service access.* For API authors, the MIME types are `application/x-blender` and `application/vnd.atlas.agent-context+json`.
* **Build and search mesh collections.** Mesh-list inputs, **Create Mesh List**, **Break Mesh List**, and concatenation nodes let workflows carry multiple meshes together. **Find Meshes by Description** selects relevant meshes from a list using their stored descriptions and a text prompt. API mesh-list input/output nodes support the same typed collections. These capabilities can also feed Blender Agent's mesh input when Blender is enabled.
* **Overlay wireframes on 3D previews.** A separate wireframe toggle and color picker let users inspect geometry over the selected preview mode. The existing exclusive wireframe mode is labeled **Just Wireframe**, and related preview controls are grouped in the top toolbar.
* **Compose 3D Scene workflows move to Blender Agent.** Saved nodes migrate to Blender Agent with GLB export enabled. Mesh and image inputs are collected into lists, existing downstream mesh-output connections remain usable, and supported model selections are retained. **Compatibility change:** audio references have no replacement input and are removed. Connected prompts may need updated `mesh_001` / `image_001` references; saved literal prompts are rewritten where possible and shortened if they exceed Blender Agent's 8,000-character limit. Migration adds a review warning and initializes a **1,000-credit maximum and 2,000-second runtime limit**. Review the prompt and limits before rerunning. For graph/API authors, `scene_composition_3d` is replaced by `blender_mcp` during graph upgrade. * **Updated Gemini model choices and defaults.** Atlas retires its remaining Gemini 2.5 choices and maps recognized saved values to the newer models. **Text + Image → Image**, including its list variant, now defaults to **Gemini 3.1 Flash Lite Image**; affected Quality Text to Image and Smart Mesh Align selections migrate too. Image-workflow migration preserves the active aspect-ratio setting and connected aspect input under the replacement field. Flash Lite Image uses Atlas's 1K tier and supports up to 14 reference images through the list variant. These are Atlas catalog changes, not an announcement that the replacement models are newly launched.
#### 3D generation * **Choose the model for Text to Origin.** The node gains a **Model** selector with **Gemini 3 Pro**, **Gemini 3.1 Flash**, and **Gemini 3.1 Flash Lite** image models. Gemini 3 Pro remains the default. The pre-run credit reservation now follows the selected model instead of a single model's cost calculation.
* **Clearer, independent 3D quality controls.** **Image → 3D** and its fallback variant now have separate **Geometry Detail** and **Texture Resolution** settings for Tripo, Trellis 2, and Titan v1. Labels and available choices match the selected engine, and applicable Tripo variants gain controls previously hidden from them. Fallback workflows can configure each selected engine separately. Existing saved choices migrate to the relevant engine fields.
#### **Workspace assets and reuse** * **Browse Blender scenes and inspect saved conversations.** Eligible users gain **Scenes** and **Conversations** asset-picker tabs. Scenes have previews; saved conversations open in a read-only viewer with the same message styling as live assistant chat, including Markdown, images, grouped tool calls, expandable details, stop reasons, and pagination. MCP's existing `list_workspace_assets` tool adds `media_type="scene"` and `media_type="agent_context"` for discovering these asset types. The picker tabs require the new visibility role; normal asset-access checks still apply.
* **Pack project assets into a ZIP, right from the project assistant.** Ask it to collect and structure project or workspace assets into a downloadable ZIP, with custom filenames and folders inside the archive, while it keeps helping with other tasks as packaging runs. A shared download list above the chat input shows progress, errors, and a Download button; downloads persist with the project, and users with project-edit access can discard them for everyone. Limits are **100 files per ZIP, 512 MiB per archive, and ten retained downloads per project**. Also available through existing MCP project-agent conversations; `pack_files_zip` is an internal assistant tool, not a new standalone MCP endpoint.
* **Select and act on multiple assets at once.** Use Select mode, Ctrl/Cmd-click, Shift-click ranges, or **Select page** in the asset picker. Download the selection as a ZIP or drag multiple assets onto the canvas together. Compatible assets are grouped into typed list inputs when available; other assets become individual input nodes. Selection applies to the current page and resets when the page or filtering context changes.
#### Image and vector generation * **Seedream 5 Pro image-layer extraction.** **Split Image into Layers** gains a Seedream engine alongside Qwen Image Layered. Use automatic decomposition or describe the elements to separate, choose 1K/1.5K/2K resolution, and select Standard or Fast prompt optimization. Results include a base image followed by up to 16 layers, plus a **Layer Metadata** JSON output with stacking order, names, descriptions, and placement bounds.
* **Revised Seedream 5 Pro sizing controls.** Image generation and editing now offer 1K, 1.5K, and 2K pixel-area tiers across aspect ratios, including very wide/tall 16:1 and 1:16 images. Matching an input image preserves its aspect ratio while applying the selected resolution tier.
#### Speech, voices, and music * **Seed TTS 2.0 joins Simple Text to Speech.** Choose Stokie, Dacey, or Tim, and provide natural-language voice instructions for tone, delivery, and pacing. Additional controls adjust speech rate, volume, and pitch.
* **Seed Audio 1.0 adds more ways to generate audio.** Generate speech, dialogue, music, or sound effects from a text prompt. Alternatively, supply up to three audio references, or one image reference with the words to speak. Speech rate, volume, and pitch controls are available here too. Audio references can be up to 30 seconds and 10 MB each; image references can be PNG, JPEG, or WebP up to 10 MB.
{% embed url="" %} * **Inworld Realtime TTS 2.0 joins Simple Text to Speech.** Choose Ashley, Dennis, Alex, or Darlene, with automatic language detection or an explicit language selection, and adjustable speaking rate. Direct delivery with bracketed instructions, non-verbal cues, and SSML pauses. The **Add Audio Tags** node also gains an Inworld target to prepare expressive speech text automatically. Inworld input is limited to 2,000 characters per generation.
* **Seed ASR 2.0 Flash joins Speech to Text.** Transcribe WAV, MP3, or Ogg recordings up to two hours and 100 MB, with 39 language choices. Options identify speakers, normalize numbers and dates, and remove filler words. A new **Transcript Annotations** output provides JSON with utterance/word timestamps and speaker labels when enabled.
* **More sound-effect engines and longer clips.** The former "Sound Effects (ElevenLabs)" node is now **Sound Effects**, with a choice of ElevenLabs SFX v2, **Stable Audio 3** for 1–120-second clips, or **Sonilo Sound Effects** for 1–180-second clips. Stable Audio also exposes a seed control. Existing nodes retain ElevenLabs as their selected engine.
* **More music and ambience generation options.** **Simple Music Generation** adds **Stable Audio 3** for 1–120-second music, **Sonilo Music** for 1–600-second music, and **Mirelo SFX 1.6** for 1–60-second ambience and sound effects. Mirelo includes an **Ambience Loop** option and an advanced **Double Output** option that repeats the generated loop. Mirelo and Stable Audio expose seed controls.
* **Trim audio with the new Audio Edit node.** Its **Simple Cut** mode extracts a clip between chosen start and end times. Set End Time to `-1` to keep everything from the start time to the end of the track; an end time beyond the track is shortened to its actual duration. This is a local audio-processing operation with a zero configured production cost.
* **Use reference audio to guide ElevenLabs music.** **Create Music (ElevenLabs)** gains two individual reference inputs and an alternative **Reference Audio List** for more references. Each 10–30-second reference guides one successive chunk of a Music v2 composition plan, with **Reference Strength** choices from Low to Extra High. Reference count cannot exceed the number of plan chunks, and reference processing is included in the cost calculation. Existing workflows gain empty reference defaults automatically.
* **Seed control for ElevenLabs composition-plan music.** **Create Music (ElevenLabs)** adds a composition-plan-only **Seed** control and a **Seed Used** output for downstream workflows.
* **ElevenLabs Music 2.5 for direct music generation.** Music v2 composition plans now render with `music_v2_5`; the direct-provider fallback for prompt-based generation also uses it. Saved legacy Music v1 section plans remain supported with their legacy model. * **New ways to transform, repair, and extend audio.** **Simple Modify Music** adds five modes: **Stable Audio 3 Transform** with adjustable transformation strength; **Stable Audio 3 Inpaint** to replace a selected time window; **Stable Audio 3 Outpaint** to generate audio before and/or after the source; **Mirelo SFX Inpaint Audio** to replace a 1–8-second sound segment in a clip up to 10 seconds; and **Mirelo SFX Extend Audio** to append sound to a 3–59-second source, up to 60 seconds total before optional loop doubling. Mirelo extension supports loopable ambience. Stable Audio edit controls cover windows up to 120 seconds, with up to 120 seconds of extension on each side for outpainting.
* **Generate music and synchronized sound effects for existing videos.** **Video Edit** adds **Mirelo SFX 1.6**, **Sonilo Video Music**, and **Sonilo Video SFX**. Mirelo accepts sound direction and a chosen output duration or "Match Input Video," up to 60 seconds. Sonilo can infer music style or sound effects from the video when no prompt is supplied; its music mode can preserve the original speech/vocals under the new soundtrack. Sonilo processes up to the first 60 seconds of a longer video.
* **Create music composition plans from visual and audio references.** The Gemini algorithm in **Composition Plan (ElevenLabs)** accepts **two reference images and two reference recordings**. Each recording can be up to **ten minutes** and smaller than **20 MB**; image references must be smaller than **7 MB**. Updated instructions prioritize the user's brief, allow more varied song structures, and give clearer guidance for vocals, instrumental passages, rhythm, and production. The default plan-generation model is now actually **Gemini 3.8 Flash** when no valid model setting is supplied; explicit platform settings remain respected. These planning references are distinct from Create Music's existing **10–30-second** conditioning clips.
#### Text models * **Updated language-model choices.** The shared model catalog adds **Seed 2.0 Mini**, **Seed 2.0 Lite**, **Seed 2.0 Code (Preview)**, and **DeepSeek V4 Pro GA**. **GLM 5.3 Flash** replaces GLM 5.2. Saved April-preview DeepSeek V4 Flash and Pro selections migrate to their GA versions; DeepSeek V4 Flash GA was already available in 0.33.1. These choices appear in compatible text/LLM nodes and model selectors. GLM and DeepSeek remain text-only in Atlas, so they are excluded from the project assistant's vision-capable selector.
* **A clearer Node Pricing page.** Search and category filters make nodes easier to find. The table shows what changes a node's price alongside its credits-per-run information, distinguishes flat rates from sampled ranges or starting amounts, and collapses free helper nodes until expanded or matched by search. Displayed ranges are guidance; the configured node's estimate or maximum before execution remains the relevant run-specific amount.
* **A refreshed workspace Members page.** A more compact table groups member identity, role, and join date, with member counts, an Invite action, and a separate pending-invitations table for copying links or revoking invites. The page displays the existing unlimited-seats entitlement where applicable; the change does not introduce a new seat allowance or change workspace permissions.
--- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation by asking a question. Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters: ``` GET https://docs.atlas.design/atlas-ai-studio-overview/release-notes/release-0.34.0.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.