Volcengine SeeDream V3

Here's the thing about SeeDream: ByteDance has never released the weights, and almost certainly never will. Seedream is the flagship of their closed, API-only image line, and this node is the only realistic way to get it running inside a ComfyUI graph. You're not loading a checkpoint - you're calling the Volcengine cloud from your workflow, paying per generation, and getting the result back as a normal image tensor. Zero VRAM, zero model downloads, and honestly the easiest "new best image model" you'll ever add to ComfyUI.

The trade is that it's a cloud call, so it's subject to keys, network, and moderation - but if your goal is prompt-faithful, polished output without babysitting a local pipeline, SeeDream V3 is a genuinely strong option. It's also a fun comparison baseline: generate the same prompt here and in local SDXL or Flux, and see how the closed model measures up. In the wider workflow it slots in anywhere a text-to-image node would: front of the graph, with the image output feeding a save, an upscaler, or an img2img chain.

How it works

The node signs each request with an AWS V4-style signature (host visual.volcengineapi.com, service cv, region cn-north-1) and POSTs your prompt plus parameters. Generation is synchronous - usually a few seconds - and the result is downloaded from a temporary URL and converted straight into a ComfyUI IMAGE tensor, so it plugs into the rest of your graph without any extra steps. Set return_url and the node also hands you the ephemeral URL as a string.

The inputs that matter

Only a handful are worth touching as a beginner:

• access_key / secret_key - your Volcengine console credentials. These are the classic access/secret pair, not the Ark API key used by the pack's video node. Different console, different key. Getting them mixed up is the #1 failure mode of this pack.
• prompt - the actual image description.
• aspect_ratio - six options (1:1, 4:3, 3:2, 16:9, 9:16, 21:9), each mapping to a ~1.5K resolution like 1536×1536 or 2016×864.
• guidance_scale - how hard it follows the prompt, 1.0–10.0, default 2.5.
• use_pre_llm - flip this on and a ByteDance LLM rewrites your prompt before generation. Nice when your prompt is short or sloppy.

seed (-1 = random) and filename_prefix round it out. Outputs are image (IMAGE) and image_url (STRING) - treat the URL as ephemeral; the tensor is what you keep.

Installing

In ComfyUI Manager, search for ComfyUI-JM-Volcengine-API and install, or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/synthetai/ComfyUI-JM-Volcengine-API
cd ComfyUI-JM-Volcengine-API
pip install -r requirements.txt

Then restart ComfyUI. The dependency list is just requests and Pillow - no torch, no model files, nothing heavy. You'll find the node under JM-Volcengine-API/Seedream.

Where people get burned

• Auth failures - the classic. Double-check the key pair and that you've enabled the visual / image generation service in your Volcengine account.
• Network - this is a China-hosted API. From outside the region it works, but expect some latency and occasional flakiness; a stable connection matters.
• Moderation - content gets screened, so prompts that would fly locally may bounce.
• The README's guidance range - it claims up to 20.0, but the node itself caps at 10.0. Trust the widget, not the doc.

It's a pay-per-call cloud model, so it won't replace local generation for free unlimited experimentation. But for polished images without the setup, it's hard to beat.