Graph Chat — the branching workspace

Response types & video controls

Every request declares what it wants back, and each modality carries its own controls:

• Image — text-to-image, instruction edits of any image node, and a prompt-adherence control that trades literal obedience for creativity.
• Video — text-to-video and image-to-video with duration choices of 5–120 seconds and 16/30/60 fps. Clips longer than five seconds are generated as chained segments, each continuing from the last frame of the one before; 30 and 60 fps are temporally interpolated. Motion and source fidelity sliders balance movement against preserving faces, clothing and composition, and video follow-ups continue from the clip's last frame.
• Text — answers from a local instruct model; asking a question about an image or video node routes through a local vision model that actually inspects the media.
• Audio — spoken answers via text-to-speech, saved as audio nodes. Any audio or video node also offers Transcribe, which creates a linked, searchable transcript node.
• 3D model — text-to-3D results land as GLB nodes that open in an interactive viewer (and in 3D Studio).

PDF document graphs

Drop a PDF onto a fresh workspace and Grafy builds a document session: it keeps the original paper, extracts the headings, summarizes the overview and each major section locally, and lays the result out as a navigable mind map beside a built-in PDF reader.

The taxonomy adapts to what you dropped — a detector reads the title, headings and section text, then applies a profile for research papers, contracts, policies, manuals, reports or proposals, with a general fallback. Category labels, colors and edge meanings are stored per node, so the legend always describes your document. Text-like attachments (.txt, .log, .md, .csv, .json) feed the request context the same way.

Merging & compare

Select multiple completed nodes and merge them into one result — as a still image or an animated MP4. Merge context is partitioned: shared ancestors are applied once, then each node's exclusive branch history is preserved separately before the final merge instruction, so a common grandparent doesn't get double-counted.

You stay in control of context everywhere: exclude specific ancestry from any request, and the exclusion is recorded on the generated node. Regenerating a leaf keeps the old version as a sibling with explicit provenance, and Compare two graphs (in the Sessions drawer) diffs whole sessions.

Prompt Studio, Batch, Auto-pilot & plugins

• Prompt Studio — templates, style packs, reusable instructions and negatives you can apply to any request.
• Batch — generate several variations in parallel from one prompt; each lands as its own node under the same parent.
• Auto-pilot — give one goal and Grafy plans a chain of steps, executing them as connected nodes you can inspect or take over at any point.
• Plugins — register your own ComfyUI workflow as a first- class tool; it then appears in the composer like any built-in modality.

Parallel jobs, progress & recovery

Multiple requests can run at once on different branches of the same session — pending nodes show a faded source preview with live progress while completed nodes stay fully interactive. Each session's jobs are independent, so you can switch sessions (or apps) while work continues, and in-flight jobs survive a page reload: reopening the app re-attaches to anything still running. Interrupted edit sessions reopen with their source canvas rather than a blank placeholder.

Sharing, collaboration & workspaces

Sessions are private to your account by default. Share one by signed-in link or invite collaborators by email as viewer or editor from the session header; the Sessions drawer separates what you own, what you've shared, and what's shared with you. The More menu adds team workspaces, project memory (facts the session should remember), API keys and webhooks for automation, and scheduled pipelines.

Exports

Any session exports as JSON (full data), Markdown, HTML, PDF, a PowerPoint deck, or a rendered storyboard MP4 — one artifact from the whole graph, downloadable from the session header. Individual nodes download their media directly.

Models & credits

The Generation settings panel selects among a central open-weight catalog (Qwen image/edit/text/vision families, Wan video, and more) plus cloud models routed through hosted providers. The list is filtered to what your hardware or plan can actually run; each option shows its runtime, whether it's cached, and its relative credit multiplier.

Charges multiply modality, resolution, duration, operation and history by model complexity — bigger models cost proportionally more, CPU runs cost less than GPU runs, and the estimate is always shown before you send. The usage dashboard breaks spending down by day, modality and model.