brennglas
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
No arguments | |||
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
Brennglas renders images and videos locally on the user's GPU; every result shows up live in the user's gallery. Work like a director when asked for a film, spot or sequence:
Plan a shot list first: per shot the subject, action, camera move, shot size and look. Keep characters, products and the setting consistent across shots by repeating the same concrete descriptions.
Create a strong start image per shot (text to image), check the preview, regenerate if it is off. For continuity, edit the previous image instead (edit mode with the earlier result as main image) or use last_frame of the previous shot. Before animating, check hands, faces, products and lettering with inspect_image: the video model carries every flaw of the start image through the whole shot. FLUX 3 Klein is weak at gripping hands; prefer relaxed poses (bag on the forearm, hand loosely at the handle) or try several seeds.
Animate each shot with image to video. LTX-2.5 (ltx25-i2v) has camera presets, look and sound; Wan Fun Camera (wan22cam-i2v) gives precise pan/zoom/roll paths but no sound. Large LTX moves need >= 5 s and pe=true or a detailed prompt. First + last frame (ltx25-flf2v) makes controlled transitions between two images.
Look at each result preview; redo a shot rather than keep a weak one. Then join the shots with concat (fade 0 for cuts, 0.3 to 0.8 for soft transitions).
Finish in concat, not in the model: one continuous music track (music, the shots' own sounds are ducked under it), the logo as a sharp overlay (logo) and an end card image (endcard). Never let the video model render logos or lettering; it smears. A dip to black (transition fadeblack) separates scenes, fade_in/fade_out open from and end on black. The user can open the result in Brennglas' editor (tab "Video editor" / "Schnitt" in the top bar) to trim shots, change single transitions and render again. Known LTX limits and what helps:
Camera presets are text appended to the prompt, not a camera path. "Tracking shot" is a sideways move; for a subject walking towards the camera use "Lead shot (subject walks towards camera)" or "Slow pull out" and say "full body in frame".
Whatever is outside the start image gets invented (extra people, props). Show the whole set in the start image.
Stage directions are suggestions ("turns to the camera" may become a turn away). When the end pose matters, generate the end image too and use first + last frame (ltx25-flf2v).
Faces in wide shots have few pixels and lose likeness: prefer medium/close shots, or 1080p for wide shots.
For an exact camera path use Wan Fun Camera (no sound, max 720p) and add sound via the music track. Recurring characters and products (same person or product in several shots):
Use an existing one from figures, or create_figure from a photo with the face clearly visible, or a packshot (any gallery id or file; make one with text to image first if needed). Name the clothing in outfit; check the returned sheet against the portrait: face shape and build must match.
Per shot generate the start image with figures=[name] and an edit workflow. For people prefer qwen21-edit: it keeps face shape and build, flux3-edit tends to widen faces and bodies. Qwen is non-commercial; for commercial work use flux3-edit (products too), where a close-up portrait helps; h3-edit also works. Prompt only the scene, aspect sets the format, up to 3 figures per scene, images add further references (a location). Do not start from the previous shot here; the figure's sheet keeps the identity, chaining edits drifts.
Check the face with inspect_image against the portrait before animating; regenerate with another seed if it differs. Medium and close shots hold the likeness best; in wide shots describe face and hair in the prompt. Call list_models first to see models, fields and option names. Prefer English prompts. Tell the user the gallery ids and links.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": true
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| list_modelsA | Lists the local image and video models with their modes, the workflow name per mode, and every input field (type, default, range, options). Call this first. Enum options can be passed by their shown name (e.g. camera "Arc left"). Models with "missing" files cannot run. |
| generateA | Starts a job in Brennglas (it appears live in the user's gallery). Pick the workflow from list_models. images: gallery ids or absolute file paths; the first is the main/start image, further ones are references (for first+last frame video, the second is the last frame). Waits up to wait_seconds, then returns the result with a preview image, or the id to pass to wait. Videos take 30 s to several minutes. Camera moves in image to video (LTX): large moves (arc, orbit, crane, drone, bullet time) only show from about 5 s; with 3 s the model falls back to a gentle push-in. For those use duration >= 5 and either pe=true or a detailed prompt that also describes the camera path. |
| figuresA | Lists the saved figures (recurring people and products) with name, kind and the gallery id of their reference sheet. Pass a name in generate.figures to put the figure into a new scene. |
| create_figureA | Saves a recurring person or product for consistent shots: renders a reference sheet (four views) from a portrait or packshot and returns its preview. The face should be clearly visible; with flux3-edit a close-up works best. Takes about 30 to 90 s. |
| waitB | Waits until a Brennglas job is done (or the timeout passes) and returns its result with a preview image. |
| enhance_promptA | Turns a short idea (any language) into a detailed English prompt for a target: "Generate image", "Edit image", "Text to video (LTX)", "Image to video (LTX)" or "Wan camera move". Optional image (gallery id or path) for image-based targets. large=true uses Qwen 3.5 9B instead of Gemma 4 E2B. Takes 1 to 4 seconds. You may also write prompts yourself; this is a helper. |
| inspect_imageA | Returns a region of a gallery image (or a video's preview frame) at full resolution, up to 1024 px. The previews from generate are only 512 px; use this to check hands, faces, products and lettering before animating. Region as fractions of the image: x, y = top-left corner, w, h = size (0 to 1). Default: the whole image. |
| last_frameA | Extracts the last frame of a finished video as a new start image, for a follow-up shot that continues seamlessly. Returns an image reference ("upload:…") to pass in generate.images. |
| concatA | Joins finished videos (gallery ids, in order) into one MP4 or WebM that appears in the gallery. Shots are matched to the first shot's size and frame rate; shots without sound get silence. fade: seconds of the transition between shots, 0 = hard cut. |
| galleryA | Lists recent results (newest first) with id, workflow, status and prompt, optionally filtered by a search text. Use the ids as images in generate, e.g. to animate an image or edit it further. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 10 tools
Each tool has a clearly distinct purpose: listing models, generating jobs, waiting on jobs, browsing results, inspecting images, extracting last frames, concatenating videos, managing figures, and enhancing prompts. The boundaries between similar areas are well described, such as figures listing versus create_figure creation, and gallery listing versus list_models.
Several tools follow a verb_noun pattern (list_models, create_figure, enhance_prompt, inspect_image), but others are noun-only or bare verbs (figures, gallery, generate, wait, concat, last_frame). The set is readable, but the conventions are mixed rather than predictable throughout.
Ten tools is well scoped for an image/video generation server, covering model discovery, generation, waiting, gallery browsing, inspection, concatenation, figure creation, and prompt enhancement. Each tool appears to earn its place without excessive surface area.
The core lifecycle is covered: discover models, generate or enhance prompts, wait for jobs, browse results, inspect images, extract last frames, concatenate videos, and manage recurring figures. Minor gaps remain around job cancellation, deleting or cleaning gallery items, and fetching a single gallery entry, but agents can work around them.