35:11
AI video generation
The video is a step-by-step walkthrough of a complete node-based Higgsfield Canvas workflow that turns a single text prompt (a turtle climbing a mountain to see the sunset) into a finished short AI video, chaining LLM, image, video, and voiceover nodes together, and argues that with the right node techniques (reference chaining, prompt structuring, connection mechanics, and cost-aware settings) this pipeline can be cloned as a reusable template for any story idea.
Higgsfield Canvas is a node-based system that keeps the interface uncluttered, letting users build a video-generation workflow step by step by chaining nodes rather than facing an overwhelming dashboard.
The entire workflow starts from one text prompt that a dual-prompt LLM node (a fixed system prompt plus a changeable user prompt) expands into five cinematic shot descriptions.
Prompt nodes are 'dumb' and require the user to supply full instructions, while LLM nodes are 'smart' and generate content themselves from those instructions.
Sequential image generation, with each new image node referencing the prior image's output, preserves character and scene consistency across frames; generating all images simultaneously breaks that consistency because each node runs independently.
Video generation happens per shot in Kling 3.0 (chosen for its claymation/cartoon style and economy), with each clip prompted using five elements: character, setting, dialogue, action, and camera movement.
When a reference image is already loaded into a video node, the character and setting description can be dropped from the prompt, leaving only action, camera movement, and dialogue if present.
Settings choices such as image resolution, standard vs. pro video mode, and sound on/off trade credits for quality or features, and the video walks through the exact credit cost of nearly every node type.
A second LLM node generates a clean voiceover script from the story concept, which is fed into a text-to-speech node (Eleven Labs) to produce narration, completing a two-stage LLM-to-TTS pipeline.
Final assets (video clips, images, audio) are downloaded individually from Canvas and assembled manually in an external editor (CapCut or Premiere Pro), with music layered underneath.
The entire node graph functions as a reusable template: cloning it and swapping only the initial idea/prompt lets a user generate an entirely new story using the same workflow.
Node-based canvas workflow — Higgsfield Canvas's core interface paradigm where discrete operations (prompt, LLM, image, video, audio nodes) are chained by linking outputs to inputs instead of using a single overwhelming dashboard. Apply: Build a story pipeline by adding nodes via right-click or the '+add node' bottom dashboard and connecting them step by step rather than trying to do everything from one screen.
Dual-prompt LLM node (system prompt + user prompt) — A structure where a fixed system prompt (instructions) and a changeable user prompt (the idea) are fed as separate external inputs into a single LLM node. Apply: Write the system prompt once as reusable instructions, then swap only the user prompt/idea for each new project so the LLM node's behavior stays consistent.
Prompt node vs. LLM node distinction — Prompt nodes are 'dumb' and simply hold text the user writes, while LLM nodes are 'smart' and generate new content (like shot descriptions) from their inputs. Apply: Use prompt nodes to store static instructions/ideas and route them into LLM nodes when you need the AI to expand or transform that text.
Sequential image generation with reference chaining — Running image nodes one at a time and connecting each generated image's output as a reference input into the next image node to maintain visual consistency. Apply: Generate image 1, connect its output to image 2's reference input, run image 2, and repeat down the chain instead of running all image nodes simultaneously.
Selective reference chaining — Connecting one sufficiently representative frame's output to multiple later frames rather than chaining every frame strictly in sequence. Apply: If frame 2 captures the needed look, connect it directly to frames 3 through 5 to save chaining steps while preserving consistency.
Clone-before-prompt node duplication — Duplicating a video or image node that already has the correct settings before writing a new prompt into it. Apply: Clone a configured node for each new shot so settings (resolution, model, mode) carry over, then just edit the prompt text on the clone.
Right-to-left node connection — Dragging a connection from an output port back to an input port (right to left) rather than left to right. Apply: When linking two nodes, drag from the output side to the input side for a more reliable connection.
White-line connection confirmation — A UI cue where a node connection line must turn white (from red) before releasing the drag, confirming a valid link. Apply: Position the cursor at the connection cross point and wait for the line to turn white before dropping it, rather than releasing immediately.
Keyboard-copy vs. right-click-copy — A distinction where copying prompt text with a keyboard shortcut (Ctrl/Cmd+C) yields clean text, while right-click 'copy' captures unwanted underlying code. Apply: Always use Ctrl+C/Cmd+C to copy prompt text between nodes instead of the right-click context menu.
Five-element video prompt framework — A prompt-writing checklist for video nodes consisting of character, setting, dialogue, action, and camera movement. Apply: When writing a video generation prompt from scratch, cover all five elements so the model has complete scene information.
Conditional prompt condensation — A rule that when an image is already loaded as a reference into a video node, character and setting description can be omitted from the prompt. Apply: With an image reference attached, write only the action and camera movement (plus dialogue if any) instead of re-describing character and setting.
Whisper Flow voice-to-text input — A third-party app that lets users dictate prompt text via a keyboard shortcut on Windows or Mac instead of typing. Apply: Trigger Whisper Flow's shortcut to speak a prompt aloud and have it transcribed directly into a Canvas prompt field.
Subjective/emotional intent-guiding language — Embedding a character's emotional state (e.g., 'struggling to keep walking') in a prompt to constrain the model's output toward the intended scene rather than describing appearance alone. Apply: Add emotional context to action descriptions in video prompts to steer the model away from unwanted behaviors like crying or giving up.
Orthogonal camera-vs-character movement specification — A prompting rule that camera movement and character movement must be described as separate instructions so they aren't conflated. Apply: Specify camera direction/speed (e.g., 'dolly left slow motion') separately from character instructions like 'keep same framing, stay steady' to avoid the model slowing the character instead of the camera.
Redundant/iterative prompting for stubborn models — Repeating the same instruction two to three times within a prompt because a model (Kling 3.0) may not comply on the first phrasing. Apply: For hard-to-enforce instructions like slow motion, restate the requirement multiple times in different phrasings within the same prompt.
Sound toggle for cost optimization — A generation setting that disables audio in video output to reduce credit cost (6 credits muted vs. 8.75 with sound). Apply: Turn off sound on video nodes when audio isn't needed for that shot to save credits.
Two-stage LLM-to-TTS voiceover pipeline — A workflow where an LLM node first generates a narration script from story constraints, then a separate text-to-speech node converts that script to audio. Apply: Feed story constraints (hook, beginning/middle/end, lesson, emotional goal) into an LLM node to draft the script, then connect its output into a voice/TTS node.
Clean-output LLM instruction constraint — A system-prompt directive telling the LLM to output only the voiceover text with no additional comments or thoughts. Apply: Instruct the script-generating LLM node explicitly to return a 'completely clean, ready-to-use reference' so its output can be piped directly into the TTS node without manual editing.
Parallel prompt queuing — Writing and queuing additional prompts into other nodes while an earlier node is still rendering. Apply: While one image or video node is generating, populate the next node's prompt so no time is lost waiting.
Template cloning/reuse — Duplicating an entire finished node workflow so it can be reused for a new story by changing only the initial idea. Apply: Clone the whole project (via 'join project' then copy to your own dashboard) and swap the opening text prompt/idea to regenerate a new story with the same pipeline.
GPT Image 2 — One of the image generation models selectable within Canvas's image node. Apply: Choose GPT Image 2 from the image node's model options when generating story frames.
Nano Banana Pro — An image generation model option where 1K resolution output costs 2 credits. Apply: Select Nano Banana Pro and choose 1K resolution for a lower-cost image generation compared to 4K (12 credits).
Kling 3.0 — A video generation model described as best suited for claymation/cartoon style and the most economical option, offering standard and pro modes. Apply: Select Kling 3.0 in the video node, set duration to 8 seconds, and pick standard mode (14 credits) or pro mode (16 credits) depending on quality needs.
Eleven Labs voice presets — The text-to-speech model used for the voiceover node, offering multiple selectable voice presets such as Talulah and Killian. Apply: Connect the generated script into an Eleven Labs voice node and pick a preset voice matching the desired narration tone.
Selectable LLM backends (Claude, Gemini, GPT) — Multiple large language models available as the engine behind Canvas's LLM node. Apply: Choose a specific LLM (e.g., Sonnet/Claude) from the LLM node's model dropdown depending on preference for generating shot descriptions or scripts.
Alternative video generation models (SeaDance, Grok, Vio) — Other video model options listed alongside Kling within Canvas's video node. Apply: Select one of these from the video node's model list as an alternative to Kling 3.0 when generating clips.
The connection-line color (red vs. white) is presented as a load-bearing UI cue: a node connection must visibly turn white before it can be dropped, or the link fails — a mechanic the video frames as 'tricky' and easy to miss.
Copy method silently determines output quality: using the keyboard shortcut (Ctrl/Cmd+C) yields clean prompt text, while right-click 'copy' pulls in unwanted underlying code, a failure mode the video explicitly warns against.
Right-to-left node connections (dragging from an output port to an input port) are described as more reliable than left-to-right, a workaround framed as learned through use rather than documented product behavior.
Kling 3.0 is characterized as 'stubborn' about slow-motion instructions, sometimes requiring the same camera-speed instruction to be repeated two to three times before the model complies.
Embedding subjective, emotional framing in a prompt (e.g., 'struggling to keep walking') functions as a soft behavioral constraint that narrows the model's action space, explicitly used to prevent unwanted deviations like the character crying or giving up.
Camera movement and character movement must be specified as separate, orthogonal instructions (e.g., 'dolly left slow motion' governs camera speed, not character speed), or the model conflates the two.
Disabling sound in video generation is framed purely as a cost lever (6 credits muted vs. 8.75 credits with sound) rather than a creative choice, revealing granular credit economics threaded through nearly every node in the pipeline.
The voiceover LLM node is explicitly instructed to output nothing but the clean voiceover text with 'no additional comments or thoughts,' a prompt-engineering move made specifically so the output can feed directly into the next node without manual cleanup.
«It's a node-based system. It's not overwhelming. You don't have all the home page stuff attacking you. It's all in here contained and you can take it step by step.»
— 00:00
«Some battles aren't won by speed. They're won by refusing to quit.»
— 00:07
«I'm going to be breaking down every detail of the workflow. But if you don't want to create it from scratch, you can actually grab a copy of it down in the description.»
— 00:55
«The system prompt is an instruction that's going to be attached at the system level, meaning your idea can change, but the system prompt will remain the same.»
— 04:52
«Don't right click and hit copy. Cuz that for some reason gets the actual code that it has.»
— 11:50
«You got to wait. It's very tricky. You got to kind of pick it where the little cross happens here. And then don't drop it until you see that line turn white, and that's when you can drop it.»
— 15:13
«Turtle stares at the sunset beyond the horizon and takes a deep breath as it begins to walk slowly towards that area.»
— 19:17
«You want your character, you want your setting. If you have any dialogue, you want that in there. You want your action and you want your camera movement. So, those five things.»
— 21:01
«Keep same framing, just so it understands to keep that framing and not like don't walk towards it, don't walk backwards, don't walk left, stay steady.»
— 26:20
«Your output will only include the voiceover, no additional comments or thoughts because I wanted the output to be a completely clean, ready-to-use reference for the voiceover node.»
— 28:08
«If I wanted to reuse the same framework, all I have to do is to change my idea.»
— 34:38
Reception
Strong positive reception with enthusiastic praise from viewers, though a technical issue was reported.
The video functions as a dense, hands-on tutorial rather than a conceptual piece, pairing a high-level node-graph philosophy with granular, tool-specific mechanics (connection gestures, copy methods, per-node credit costs) that make it a practical reference for replicating the exact workflow shown.
35:11