Banana Slides: an AI-native PPT generator that turns prompts into editable decks

What Banana Slides actually automates

Banana Slides addresses the gap between AI text generation and usable presentation files. Many AI slide tools return a rendered preview or a PDF, which is useless if you need to adjust the deck afterward. This project targets that problem by producing editable PPTX output from several input types: a single sentence, an outline, a page description, or uploaded source material such as documents and images. The README also describes template control, meaning you can upload an arbitrary template image and have the AI parse its style. The intended user is someone who wants a starting deck in minutes, not a finished design product, and who is willing to edit and refine the result rather than present it unchanged. The project name references nano banana pro, which appears to be the underlying model family, though the README does not explain what that model is or how it is hosted.

The generation pipeline: from description to editable file

The repository does not include a detailed architecture document in the supplied material, but the release notes and feature descriptions reveal the main stages. You begin with a description, outline, or uploaded material. The system then generates an outline, and the release notes for RC3 mention outline stream isolation, which suggests outlines are produced incrementally rather than in one blocking call. After the outline, the system builds slide content and matches each page to a template. RC3 also introduced template matching and field contract v2, which implies a structured data flow between the language model output and the rendering layer. The editable PPTX export was improved in the same release. Image generation is a separate track: RC7 notes that the internal main model was updated from gpt-5.4 to gpt-5.6-terra while the image model remains gpt-image-2, and RC6 added support for SenseNova U1 image generation. This separation matters because a failure in the image step, such as the Codex OAuth 400 error fixed in RC7, can break an otherwise valid generation.

Conversational editing and template matching

Two features distinguish Banana Slides from a simple prompt-to-PPTX wrapper. The first is conversational editing, described as making changes to specified regions by voice or text. The README's Chinese description says you can verbally request modifications to a specific area, which implies a mapping between natural language edit requests and the slide layout structure. The second is per-page template matching, added in June 2026. You can use a single template for all slides or assign a different template to each page. Templates can be uploaded as images or PDFs, and the AI parses the template style automatically. The system can then match a template to each page automatically, or you can bind templates manually. The two modes are switchable at any time, which is a practical design choice because automatic matching will not always pick the right layout for a dense data slide or a title page.

Deployment: Docker, desktop, and the public demo

The README points to a local Docker deployment path and a desktop release candidate. The latest desktop version is v0.9.0-rc.7, and release notes describe installer behavior such as update checks, download progress, failure retry, and restart-to-install. The public demo at bananaslides.online runs with fixed model configurations from Inferera, APIMart, and Volcano Agent Plan, with API keys isolated per visitor. The public demo does not keep history, so the README advises saving the preview page link to return to a deck. Site owners can set PUBLIC_DEMO_ADMIN_PASSWORD in the .env file to view history through the /admin/history route. This is a meaningful operational detail: history is not a default feature, and the public demo is deliberately stateless. For a self-hosted instance, the exact commands to start the Docker container are not present in the supplied material, so you would need to consult the documentation site at docs.bananaslides.online for the full install steps.

Model provider dependencies and configuration friction

Banana Slides does not ship with its own model. It depends on external providers, and the release history shows this is a recurring source of breakage. RC4 fixed LazyLLM online providers being unavailable in the desktop build, including qwen, and added missing SOCKS proxy dependencies. The same release notes mention that the packaged desktop version now registers 11 LazyLLM online providers: qwen, doubao, deepseek, glm, kimi, minimax, sensenova, siliconflow, ppio, aiping, and openai. RC6 fixed APIMart OpenAI-compatible async image tasks, non-streaming requests, and 1K/2K/4K resolution passing. RC7 fixed a Codex OAuth issue where image generation returned 400 because the internal main model was outdated. The pattern is clear: every provider has quirks, and this project absorbs them through release fixes. If you plan to self-host, you should expect to configure provider credentials carefully and to track release notes for provider-specific patches. The README also mentions MinerU credential errors for PPT transformation, which suggests document parsing is a separate service with its own credentials.

Licensing, maturity, and maintenance cost

The project is licensed under AGPL-3.0, which has real consequences for adoption. If you modify the code and offer it as a network service, you must make your modified source available to users of that service. This is a stronger obligation than MIT or Apache-2.0, and it rules out internal proprietary forks unless you keep them strictly internal and do not serve them over a network. The project is also not at a stable release. The most recent tag is v0.9.0-rc.7, and the release cadence is roughly weekly or biweekly during August and September 2026. Maintenance cost is therefore ongoing: each release fixes provider-specific issues, and the desktop packaging has its own update mechanism. The last push was September 9, 2026, so the project is actively maintained. The README does not state a long-term support policy or a migration path between release candidates, so you should budget time for tracking releases if you deploy it.

Where it falls short and what to compare it with

The most obvious limitation is that the supplied material contains no evidence of offline or local model support. Everything points to external APIs, which means generation cost, latency, and network dependency are unavoidable. The public demo is stateless, and even self-hosted history requires an admin password route rather than being a default feature. The project is also oriented toward a specific workflow: generate, then edit conversationally. If your need is a one-shot polished deck with no further edits, a tool that outputs a designed PDF or an image-based slide might serve you better. The closest alternative approach is a traditional template-based generator like a slide builder that fills predefined layouts without an LLM, or a different AI slide tool that does not expose editable PPTX. The actual difference is in the editing loop: Banana Slides tries to make the AI part of an iterative design conversation, while template-only tools treat generation as a one-way operation. The README's phrase about oral modification of specified regions is the core bet of this project, and it is also the hardest part to get right, since it requires the model to understand both the slide structure and the user's spatial reference.

Editorial conclusion

Adopt Banana Slides if you need a self-hostable, AI-native pipeline that turns prompts, outlines, documents, or template images into editable PPTX files, and if you can accept the AGPL-3.0 license and the cost of running your own model endpoints. Do not adopt it if you require a stable, non-rc release, if you want a fully managed SaaS with history retention, or if your team cannot handle the operational burden of configuring multiple AI providers. Before committing, verify which model provider you will use, confirm that the image generation endpoint supports the async and resolution parameters the project expects, and test the PPTX export path on your target operating system, since release notes show platform-specific fixes for Windows and desktop packaging. The project is in release candidate state as of v0.9.0-rc.7, so treat production use as early adoption with expected edge cases.

Official sources

1. Anionex/banana-slides on GitHub
2. License: AGPL-3.0
3. Project website
4. README
5. Releases