Ideogram 4's acceleration update ships two single-model graphs, FAST and Instant, both INT8-quantized, both 12 nodes, and — this is the part that matters — both drive the model through the Ideogram4PromptBuilderKJ structured prompt builder. This comparison is parsed from the two JSON files shipped with the episode.
What the two graphs actually contain
| FAST | Instant | |
|---|---|---|
| UNET | Ideogram4-fast_int8-convrot-simple.safetensors |
Ideogram4-instant_int8-convrot-simple.safetensors |
| Text encoder |
qwen3vl_8b_fp8_scaled.safetensors (type ideogram4) |
same |
| VAE | flux2-vae.safetensors |
same |
| Prompt path |
Ideogram4PromptBuilderKJ (enabled, mode 0) |
Ideogram4PromptBuilderKJ (enabled, mode 0) |
| Sampler |
euler / normal, 20 steps, CFG 1 |
euler / normal, 8 steps, CFG 1 |
| Latent |
EmptySD3LatentImage 2048 × 880
|
EmptySD3LatentImage 2048 × 880
|
| Nodes | 12 | 12 |
Both graphs' CLIPTextEncode nodes are empty: the prompt comes from the builder, not from a text box. That is a real difference from a typical text-to-image graph, and it means the prompt you write lives in the builder's fields.
What the builder does in practice
The builder in both files carries a full natural-language prompt (the sample is a Y2K cyber-fashion shot) plus the 2048×880 target size. In other words it is doing prompt-shaping and sizing together; the sampler then renders that conditioning.
FAST vs Instant — the real trade-off
- FAST, 20 steps: more sampling budget for the structured prompt. Use it when the prompt is detailed and you want the graph's full conditioning honoured.
- Instant, 8 steps: same builder, same encoder, same latent, a quarter of the steps. Use it to explore prompt directions, then move the winner to FAST.
Because both use CFG 1, neither has a negative-prompt stage — the choice is steps, not guidance.
How to run
- Open the graph on RunningHub (invite code
rh-v1111): FAST https://www.runninghub.ai/post/2078065107753590786/?inviteCode=rh-v1111 · Instant https://www.runninghub.ai/post/2077705321124028417/?inviteCode=rh-v1111 - Fill the
Ideogram4PromptBuilderKJfields (prompt + target size). - Run; the graph saves the image and shows a preview.
Limits
- INT8 weights are a speed/memory trade-off the episode adopts deliberately; they are not a quality guarantee.
- 20/8 steps are the shipped settings, not universal optima.
- The 2048×880 latent is the shipped default; change it for other aspect ratios.
- CFG 1 means no negative-prompt lever in either graph.
- The prompt builder is a community node; its fields may change between versions.
Episode resources
- RunningHub FAST: https://www.runninghub.ai/post/2078065107753590786/?inviteCode=rh-v1111
- RunningHub Instant: https://www.runninghub.ai/post/2077705321124028417/?inviteCode=rh-v1111
- YouTube: https://www.youtube.com/watch?v=g7HXab6m4Ys | Bilibili: https://www.bilibili.com/video/BV1BJNd6vE27
Optional API platform - 88API (affiliate link)
Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.
https://88api.ai/sign-up?aff=usx3
Disclosure: I may earn a commission from eligible purchases after you register through this referral link.
Top comments (0)