CogView4
CogView4 is THUDM’s 6B-parameter diffusion-transformer text-to-image model. It uses a GLM language model as its text encoder and a 16-channel VAE. InvokeAI supports the base CogView4-6B checkpoint. CogView4 support is basic: plain generation and Canvas image editing work, but the adapters and controls that other families offer are not available.
License
Section titled “License”CogView4-6B is released under Apache 2.0, which allows commercial use. The download is not gated.
This is a summary, not legal advice. Read the full license on the model’s HuggingFace page before using a model commercially.
Hardware
Section titled “Hardware”The Diffusers download is about 31 GB, including the GLM text encoder and VAE. See System Requirements for general guidance, and the Low VRAM guide if you run out of memory.
Installing
Section titled “Installing”Install CogView4 from the Model Manager’s starter models. Its source is THUDM/CogView4-6B. This
family has no starter bundle.
CogView4 installs in Diffusers format only, as a single folder that contains the transformer, GLM text encoder, tokenizer and VAE. No other models are needed, and the components cannot be replaced individually. Single-file and GGUF CogView4 models are not recognized.
Generation settings
Section titled “Generation settings”When you select CogView4, InvokeAI applies the defaults from the model’s reference example:
| Setting | Default |
|---|---|
| Steps | 50 |
| CFG Scale | 3.5 |
| Resolution | 1024 × 1024 |
- CFG Scale applies true classifier-free guidance.
- Width and height must be multiples of 32.
- Negative prompts are supported and always used.
- No scheduler choice. CogView4 uses its own flow-matching sampler, so the Scheduler control is hidden.
- Prompt weighting (Compel syntax) does not apply to CogView4. See Prompt Syntax. The GLM encoder reads up to 1024 tokens, so long, descriptive prompts are fine.
Canvas
Section titled “Canvas”On the Canvas, CogView4 supports text-to-image, image-to-image, inpainting and outpainting. LoRAs, ControlNet, reference images, regional guidance and PiD decoding are not available for CogView4.
Workflow editor
Section titled “Workflow editor”The CogView4 nodes are Main Model - CogView4, Prompt - CogView4, Denoise - CogView4, Image to Latents - CogView4 and Latents to Image - CogView4. Each prompt needs its own Prompt - CogView4 node. Connect them to the denoise node’s positive and negative conditioning inputs. Both are required.