Deepy

Deepy is WanGP's conversational media assistant. It can generate, inspect, edit, extract, transcribe, merge, and transform images, video, and audio while keeping conversation context.

Deepy comes in two versions:

• Deepy Zero is the lightweight, fast version for straightforward requests. It should spent less time thinking and is designed to work well with a smaller LLM (for instance Qwen3.5 4B or 9B).
• Deepy Prime is the advanced version for ambitious, multi-step work. It can discover available models and capabilities, plan connected actions, combine several image, video, and audio assets, and use external MCP services when configured. It requires the Qwen3.8 VL 27B model.

Both versions share the same chat, default deepy settings and templates, Gallery integration, VRAM policy, context window, compaction, and interruption handling.

This guide covers:

• general guidelines
• enabling Deepy
• configuring Deepy in the web UI
• linking WanGP settings files to Deepy generation tools
• using selected and previous media naturally
• understanding which generation settings Deepy can override directly
• asking Deepy about available LoRAs and current defaults
• using Deepy from the CLI

Deepy can make mistakes, so verify important results.

General Guidelines

Once enabled, open the Deepy chat window by clicking Ask Deepy in the left dock.

Both versions can generate images, video, and audio. Deepy Zero is best for focused requests using your selected tools and templates. Deepy Prime can plan longer workflows and combine several generated or existing media assets. Content produced by either version appears in the Image / Video Gallery or Audio Gallery at the top right of WanGP's Media Generator tab.

Deepy can also work with User Imported Media:

1. Expand the section Media Info / Late Post Processing / Import Media
2. Switch to the Import Media to Galleries tab
3. Select files to Import
4. Click Import Videos / Images / Audio Files

Once the media are in the galleries, you can refer to them using wording like the last audio file, the selected video or describe their content (Deepy will query the prompts stored in the generation metadata if they exist).

Deepy relies on predefined Template Settings for its seven generation tasks (Generate Image, Generate Video, Edit Image, Generate Video with Speaker, Generate Song, Generate Audio from Description, and Generate Audio from Sample). Deepy Zero directly uses the curated template assigned to each tool. Deepy Prime starts from the corresponding default template when no model is named, but can discover and compare other compatible models when the request requires it.

For Deepy Prime, every derived generation step follows the same rule. If a video workflow needs a master image and edited end frames, Prime uses the current default Generate Image and Edit Image templates directly instead of browsing models first. It performs model discovery only when the user asks to choose or compare models, names a model, or a required capability is demonstrably incompatible with the default template. Template settings already include model defaults, so fetching raw model defaults afterward is unnecessary.

WanGP comes with builtin templates ready to use but you may as well link presaved settings. You can access Deepy settings by clicking the Settings control on the right of the Deepy chat window.

In the web UI, Deepy settings changes take effect for the current Deepy session as soon as you make them. Click Save Deepy Settings at the bottom of the settings panel when you want to write those settings to disk for future WanGP sessions.

You can also define default width, height, frame count, audio duration, and seed in the Deepy Settings window. Select Use by Default Always Dimensions / Durations / Seed Below to apply them instead of the corresponding template properties without editing the templates.

You can also ask Deepy directly to override supported template settings such as width, height, frame count, audio duration, FPS, LoRAs, or inference steps.

Enabling Deepy

Deepy is available only when these base conditions are met:

1. Deepy is set to Deepy Zero or Deepy Prime rather than Disabled.
2. Prompt Enhancer is set to a supported Qwen3.5VL mode.

Deepy Prime additionally requires Compaction Type When Cache is Full to be set to Summarize and Context Window Tokens to be at least 32,000. Selecting Deepy Prime in the Configuration UI automatically raises a smaller context to 32,000 and selects Summarize. Configuration saving and runtime startup both reject an invalid Prime configuration.

Open the Configuration plugin and go to the Prompt Enhancer / Deepy tab.

Required Prompt Enhancer modes:

• Qwen3.5VL Abliterated 4B
• Qwen3.5VL Abliterated 9B

Deepy settings in that tab:

• Model used to power Prompt Enhancer / Deepy: selects the shared language/vision model.
• Speculative Decoding: Auto enables it for Qwen3.5 9B with at least 12 GB VRAM or Qwen3.8 27B with at least 24 GB VRAM. Explicit Yes and No remain available.
• Deepy: selects Disabled, Deepy Zero for lightweight, straightforward work with curated WanGP tools and templates, or Deepy Prime (requires Qwen3.8 VL 27B LLM) for advanced planning, model discovery, multimedia workflows, and optional external MCP capabilities. Selecting Prime raises the context window to at least 32,000 tokens and selects Summarize compaction.
• Allow Deepy to Read the Filesystem: disabled by default. When enabled, both Deepy Zero and Deepy Prime can list directories, inspect files, and use existing media paths in tool inputs. When disabled, Deepy must use Gallery/media ids and the filesystem tools are not exposed.
• Deepy VRAM Loading Mode: controls whether Deepy stays in VRAM, unloads when idle, or unloads only when another WanGP component needs VRAM. The more Deepy stays in VRAM, the more responsive.
• Context Window Tokens: how much conversation and tool history Deepy tries to keep live
• KV Cache Quantization: Auto enables fast INT8 KV cache when GGUF kernels 1.0.11 or newer are installed; explicit BF16 and INT8 choices remain available.
• Compaction Type When Cache is Full: either discard the oldest conversation entries at the limit or summarize older completed turns and completed action groups at the lower of 85% context usage and 4,096 tokens before the KV-cache limit. Summarize requires at least 32,000 context tokens and generates directly from synchronized KV memory, appending only a short internal instruction instead of replaying serialized history. The summary is durable working memory: it preserves goal-critical findings and decisions, clearly separates completed work from remaining work, and instructs Deepy to trust successful results instead of repeating completed actions. During a long active turn, summarization runs only at a safe boundary after a complete tool result and before the next model decode; it preserves the original user request and two newest assistant/tool action groups verbatim while summarizing older work. Unfinished decoded output is never summarized. During summarization or current-turn space recovery, the status bar is the only transient UI notification. Only a successfully committed summary creates a chat entry: an expandable Context block at its chronological position inside the active turn. Verbose level 2 also prints the complete summary to the console. In Summarize mode, reaching one decoder segment's max_tokens is never considered the end of the turn. If another complete segment fits, Deepy continues immediately from live KV without compaction, replay, or a summary entry, preserving the same unfinished Thought block. It shows Making room to continue... and rebuilds context only when another full segment cannot fit. If summarization genuinely fails, Deepy progressively removes older assistant/tool steps while retaining the newest two; dropping a whole turn is the final fallback. Whenever complete turns must be deleted, Deepy retains a compact system record of the five most recent removed user requests, truncating each to 256 characters and stating that their corresponding answers were removed. This record is discarded only if the active request otherwise cannot fit.
• Deepy Zero Prompt: edits independent extra instructions for Deepy Zero.
• Deepy Prime Guidance: edits standing user guidance directly. It is prefilled to prefer the highest-quality base or full model unless the user prioritizes speed or names another model, and is appended to Deepy Prime's trusted system instructions rather than replacing them.
• External MCP Servers (JSON): optional stdio, SSE, or Streamable HTTP MCP server definitions used only by Deepy Prime. External tool names are prefixed with their server name to avoid collisions.

When the requirement is met, the Ask Deepy launcher appears in the WanGP web UI.

Deepy Prime always includes WanGP's in-process MCP server and automatically loads its trusted wangp_agent operating guide. Its WanGP generation calls use the active browser session's Gradio queue, so jobs and generated media remain visible in the normal queue and Galleries while Deepy stays unloaded from VRAM. External servers extend that tool set; their prompts are not loaded automatically.

Because Deepy Prime runs inside the same WanGP process, its in-process MCP definition intentionally omits remote Gallery upload/download tools. It can reuse Gallery ids directly, and when filesystem access is enabled it can also use local paths. Gallery ids observed by the MCP session remain resolvable while their files exist after Keep Previous Generations in Gallery trims their visible UI rows; remembered records remain discoverable with in_gallery: false without being reinserted into Gradio. List Files optionally filters extensions and returns names, paths, and byte sizes. Query File accepts a Gallery/media id or path and returns resolution, frame count, FPS, duration and audio-track information for visual media; duration, sample rate, channels and track count for audio; or UTF-8 text up to 16,000 characters.

Deepy Prime receives the current Deepy Settings > General Properties precedence as system-level guidance at session startup. With Use by Default Dimensions / Durations / Seed defined in Templates Settings Used, inactive panel defaults are not injected or returned, and the template query keeps those template-owned fields in settings. With Use by Default Always Dimensions / Durations / Seed Below, the standing defaults relevant to the requested tool are injected and returned separately as general_properties, while conflicting resolution, frame-count, duration, and seed keys are removed from template settings. Explicit user overrides still take precedence. Changes to these General Properties change the system-prompt signature and are injected on the next turn.

External server configuration example:

{
  "filesystem": {
    "transport": "stdio",
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-filesystem", "C:\\media"]
  },
  "remote": {
    "transport": "streamable-http",
    "url": "http://127.0.0.1:9000/mcp"
  }
}

Deepy Web Settings

Open Ask Deepy, then open the Settings panel.

The settings panel contains two expanded sections:

• Generation Properties
• Template Settings used by Tools

All changes in this panel are used immediately by the current Deepy web session. To keep them for future sessions, click Save Deepy Settings at the bottom of the panel.

Generation Properties

• Auto-abort or remove Deepy-started generation on Stop/Reset.
  Controls whether Deepy-created queue work is cancelled or removed when you stop/reset Deepy.

• Default Dimensions / Durations / Seed Choose whether Deepy uses these properties from each selected template or always replaces them with the panel defaults below.

• Width and Height
  Default size overrides used only when template properties are disabled.

• Default Number of Frames and Default Audio Duration Default video length in frames and audio/song duration in seconds when template properties are disabled. Audio duration defaults to 10 seconds.

• Number of Frames
  Default frame-count override for Generate Video, used only when template properties are disabled.

• Seed (-1 for random)
  Default seed override, used only when template properties are disabled. -1 means random.

Inference steps, FPS, LoRAs, and other model-specific values remain template-driven unless you ask for one of the supported per-request overrides described later in this guide.

Tool Templates

Deepy has seven generation-tool template selectors:

• Media Generator
• Video With Speech
• Image Generator
• Image Editor
• Song Generator
• Speech From Description
• Speech From Sample

Each row has:

• a dropdown that selects the current template for that tool
• + to link that tool to the currently selected WanGP user settings file (in the dropdown in the upper left part of video gen tab )
• trash to remove the current live link and go back to the previous or default template

Changing a template selector updates the active Deepy web session immediately. Click Save Deepy Settings if you want to reuse the same selectors the next time you launch WanGP.

Deepy shows the selected template in the chat transcript for generation tools, for example:

Generate Image [Z Image Turbo]
Generate Video [LTX-2 2.3 Distilled 1.0]
Edit Image [Flux Klein 9B]

Save Deepy Settings

Click Save Deepy Settings at the bottom of the Deepy settings panel to persist the current web settings to disk.

That save includes:

• generation-property values such as auto-abort behavior, template-property usage, width, height, frame count, audio duration, and seed
• the currently selected Deepy template for each generation tool

Linking WanGP Settings to Deepy Tools

Deepy templates are either:

• built-in Deepy templates shipped with WanGP
• live links to WanGP user settings files

Link a tool from the UI

Practical workflow:

1. configure a normal WanGP generation the way you want
2. save it as a WanGP user settings file
3. select that user settings JSON in WanGP's Lora / Settings dropdown
4. open Deepy settings
5. click + next to the Deepy tool you want to link
6. confirm the link

When you use the tool later, Deepy reads the linked WanGP settings file directly, so changes to that file are picked up automatically.

Important behavior

• Only WanGP user settings selected from the Lora / Settings dropdown can be linked this way.
• System profiles and LoRA presets are rejected.
• If the linked WanGP settings file changes later, Deepy sees the updated content automatically.
• If the linked file disappears, Deepy falls back to that tool's default template.
• If the linked file still exists but is no longer eligible for that tool, the tool returns an eligibility error.
• Built-in templates cannot be deleted from the UI.
• Linked templates are the right place for model-specific settings that Deepy does not expose directly. Deepy can still override width, height, frame count, audio duration, FPS, inference steps, and LoRAs on the supported tools.

How Deepy Interprets Media References

Deepy is designed to let you refer to existing media naturally.

In practice, Deepy will usually:

• prefer the currently selected image, video, or audio item when you say selected, current, this image, this video, this audio, or this frame
• use the selected video's current playback time when you refer to the selected frame or the current frame
• resolve short references such as last image, previous video, or last audio
• resolve older outputs when you describe a previous result
• ask for clarification instead of inventing a result when a reference is ambiguous

You can still use internal media ids such as image_1 or video_3, but usually you do not need to.

Using Selected Media

In the web UI

For an image:

1. click the image you want
2. ask Deepy something like:
   • edit this image so the sky is stormy
   • inspect the selected image and tell me whether the hands look correct
   • use the selected image as the start frame for a short video
   • use this image and the last audio clip to make a talking video

For a video:

1. select the video
2. scrub the player to the moment you care about
3. ask Deepy something like:
   • inspect this frame and tell me whether the face is sharp
   • extract the selected frame as an image
   • cut a 3 second clip starting at the selected time
   • transcribe this video
   • mute this video
   • replace the audio of the selected video with the last extracted audio

For audio:

1. select or import an audio file
2. ask Deepy something like:
   • transcribe this audio
   • transcribe this audio with word timestamps
   • create speech from this sample saying: Welcome to WanGP
   • use this audio with the selected image to make a talking video

If your voice sample is inside a video, Deepy can extract the audio first.

Previous outputs

Deepy can also resolve references such as:

• last image
• previous video
• last audio
• the robot dancing image
• image_2
• video_3

What You Can Ask Deepy To Do

Deepy Zero is intended for direct requests such as generating one asset with a selected template, editing or animating selected media, extracting a clip, resizing a file, or producing a transcript. Deepy Prime can perform the same work while also selecting among compatible models and coordinating several dependent actions across multiple media.

• generate images, edit images, generate videos, generate talking videos from a still image plus speech audio, and create speech audio from a voice description or a voice sample
• create solid-color frames for transitions, blank frames, or color cards
• inspect images and video frames, and read local image, video, or audio details such as dimensions, duration, FPS, frame count, or audio track count
• extract images, video clips, or audio clips; transcribe audio or video; mute videos; replace audio; resize/crop media; and merge videos
• tell you which LoRAs are available for the current generation tool and which defaults a generation tool will use right now
• answer WanGP-specific usage questions by searching the bundled docs

Audio Transcription

Deepy can transcribe either audio or video.

• Segment timestamps are returned by default.
• Ask for word timestamps if you need more detailed timing.
• If a source has multiple audio tracks, mention which track you want.

Example requests:

Transcribe the selected video.

Transcribe audio track 2 from the selected video.

Extract the video excerpt that starts with 'I will be back'.

Example Requests

Generate a cinematic image of a robot violinist on a rainy Paris rooftop at night.

Edit the selected image so the background becomes a neon alley while keeping the character identity, and use 8 inference steps.

Generate a short video of a paper boat floating through a glowing cave river at 24 fps with 97 frames and 8 inference steps.

Generate a video of a dog playing under the rain using the Lego lora

Use the selected portrait and the last audio clip to make a talking video.

Create speech from this sample saying: Welcome to WanGP.

How do I use VACE for outpainting?

Multi-step requests are where Deepy Prime is most useful:

1) Generate an image of a robot disco dancing on top of a horse in a nightclub.
2) Edit the image so the setting stays the same, but the robot has gotten off the horse and the horse is standing next to the robot.
3) Verify that the edited image matches the description; if it does not, generate another one.
4) Generate a transition between the two images.

Create a high quality portrait that represents you well. Then create a speech sample in which you introduce your capabilities. When done generate a talking video from the portrait and the generated speech.

Deepy CLI Mode

Launch Deepy in CLI mode with:

python wgp.py --ask-deepy

At startup, the CLI prints the Deepy logo and preloads the prompt-enhancer runtime so Deepy is ready before the first prompt.

Prompt entry

Interactive multiline entry:

• Enter: send the current prompt
• Ctrl+Enter: insert a newline on terminals that expose it
• Alt+Enter: insert a newline
• Ctrl+J: newline fallback
• Ctrl+S: stop the current Deepy turn while it is running
• Shift+Enter: not available here because the console reports it as plain Enter

CLI media selection

The CLI has its own virtual gallery. Add files to it, select one, and optionally set a playback time or frame for the selected video.

Examples:

/video E:\media\my_clip.mp4
/frame 120
inspect the selected frame and tell me whether the subject is centered

/audio E:\media\voice.wav
transcribe the selected audio with word timestamps

When a Deepy tool generates media in CLI mode, the CLI prints the generated output path.

CLI commands

Media:

• /add <path>: add and select an image, video, or audio file
• /image <path>: add and select an image file
• /video <path>: add and select a video file
• /audio <path>: add and select an audio file
• /list [scope]: list known media; scope can be all, media, image, video, or audio
• /media [scope]: alias for /list
• /clear-media: remove all virtual gallery media

Selection:

• /select <ref>: select media by id, list index, or name fragment
• /select-video <media_id>: select a video by media id
• /selected: show the currently selected media
• /selected-video: show the selected video media id
• /time <secs>: set the selected video's playback time
• /frame [index]: show or set the selected video frame, 0-based

Deepy settings:

• /settings: show the current CLI Deepy settings
• /size [WxH]: show or set default generation size and disable template properties
• /frames [count]: show or set default gen_video frame count and disable template properties
• /duration [seconds]: show or set default audio duration and disable template properties
• /seed [value]: show or set the default generation seed and disable template properties
• /template <tool> <variant>: set the template for any Deepy generation tool
• /templates [tool]: list available template variants
• /template-props [on|off]: show or toggle whether Deepy uses resolution, frame, audio-duration, and seed properties from templates

Session:

• /help: print the CLI command summary
• /reset: clear the Deepy conversation but keep the virtual gallery media
• /quit: exit the CLI session

Examples:

/template gen_image "Z Image Turbo"
/template gen_video "LTX-2 2.3 Distilled 1.0"
/size 1280x720
/frames 97
/seed -1

Practical Tips

• Deepy works best when your request clearly states the goal and how current media should be reused.
• Use Deepy Zero for focused requests and Deepy Prime when the result requires planning, model selection, or several connected media operations.
• Deepy Prime can infer a workflow from the requested outcome, but listing mandatory steps or constraints explicitly makes the result more predictable.
• If you need a model-specific setting that Deepy cannot override directly, store it in the linked template.
• Ask Deepy for available LoRAs or current defaults when you switch templates and want to confirm the setup.
• For image and video requests, be explicit about any must-keep details such as subject identity, composition, or mood.
• If you want Deepy to use the current video moment, scrub the selected video first, then refer to this frame or the selected frame.
• For transcription, mention if you want word timestamps or a specific audio track.
• If a tool fails, Deepy will tell you rather than inventing a result.
• Stopping a turn immediately requests cancellation of any active Deepy Prime MCP generation job. Deepy preserves every fully completed message and assistant/tool group exactly, discards only an unfinished trailing decode fragment or incomplete tool group, and records that the request was interrupted so it is not resumed automatically. It retains the latest compatible safe KV boundary; messages completed after that boundary are appended on the next request. Normal summarization handles the preserved turn if it later approaches the context threshold, and selection or deletion occurs only when the complete state genuinely cannot fit.
• For WanGP-specific questions, you can ask Deepy directly instead of searching the docs manually.
• Install GGUF kernels for fast inference and low VRAM.