How can I improve quality without huge render times?

Start at 20 steps for drafts. For finals, raise to 30–40 steps, keep 720p base resolution, and enable HunyuanVideo15SuperResolution. If you need 1080p, use the latent upscaler path but expect longer runtimes.

I get OOM errors or black frames. What should I change?

Lower width/height and/or frame count, set weight_dtype to fp8_e4n3fn, switch to VAEDecodeTiled, and disable the 1080p upscaler group. Close other GPU apps and reduce CFG or steps if needed.

How do I control camera movements and emotion in the video?

Describe shots explicitly in the prompt (e.g., “handheld close-up, slow push-in, rack focus, emotional smile”). Mention lens and lighting cues (“35mm, golden hour, volumetric light”). Keep sentences concise and avoid conflicting directives.

What does EasyCache do, and when should I use it?

EasyCache stores intermediate encodings to speed up iterations (useful when only small prompt tweaks change). It trades some visual fidelity for speed; use it for look development, then disable for final renders.

Hunyuan Video 1.5 Text to Video

This workflow turns text prompts into coherent 720p videos using the Hunyuan Video 1.5 model, with optional 1080p upscaling and super-resolution. It’s organized into clear groups: Step 1 - Load models, Step 3 - Prompt, Step 4 - Video Size, Custom Sampler, and an optional Video Upscale 1080P (Ctrl-B to enable). The pipeline loads Hunyuan’s text encoders with DualCLIPLoader and encodes your prompt via CLIPTextEncode. EmptyHunyuanVideo15Latent defines the temporal canvas (resolution, frame count, FPS), while RandomNoise sets the seed. UNETLoader brings in the Hunyuan Video 1.5 UNet, and SamplerCustomAdvanced (driven by KSamplerSelect, BasicScheduler, and CFGGuider) performs diffusion steps to synthesize video latents at 720p. VAELoader plus VAEDecode (or VAEDecodeTiled for memory efficiency) converts latents to frames, which CreateVideo stitches into a final clip, saved with SaveVideo.

For higher fidelity, the workflow includes a latent-space upscaling path. LatentUpscaleModelLoader and HunyuanVideo15LatentUpscaleWithModel can boost resolution before decoding, and HunyuanVideo15SuperResolution refines details after generation. EasyCache is available to accelerate repeated runs by caching encodings, at the cost of some visual quality. The default step count is set to 20 for reasonable render times; increasing steps (e.g., toward the Hunyuan team’s original settings) improves detail but extends inference time. If you encounter VRAM limits, set weight_dtype to fp8_e4n3fn and switch to VAEDecodeTiled.

API

Use this workflow from code

Every Comfy workflow is a JSON graph. The payload below is this workflow, exactly as ComfyUI runs it — fetch it from the URL, keep it in version control, or load it in ComfyUI and run it node by node.

video_hunyuan_video_1.5_720p_t2v.json

Fetching workflow JSON…

Run it from TypeScript or Python with the Comfy SDK. The same code targets Comfy Cloud or a ComfyUI you host yourself — only the base URL changes.

// Install (beta)
npm i @comfyorg/sdk

// Run "Hunyuan Video 1.5 Text to Video" (TypeScript)
import { Comfy } from "@comfyorg/sdk";

const client = new Comfy({ apiKey: "comfyui-..." });

// This workflow, exported in API format (see note below)
const wf = await client.workflows.fromFile("video_hunyuan_video_1.5_720p_t2v_api.json");

wf.setInput("44", "text", "your prompt here");  // CLIPTextEncode

const job = await client.run(wf);
await job.getOutputs("102")[0].toFile("output.png");  // SaveVideo

The SDK takes a workflow in API format: open this workflow in ComfyUI and use File → Export Workflow (API).

Comfy Cloud API access requires a plan with an API key.

SDK docs Get an API key

FAQ

Frequently Asked Questions