ComfyUI Node
Eric ERNIE-Image UltraGen
Two-stage hierarchical generation. Returns final image + Stage 1 composition preview.
Eric ERNIE-Image UltraGen
- pipeline
- prompt_embeds
- s1_negative_embeds
- s2_negative_embeds
- final_image
- stage1_preview
◄prompt►
◄resolution5:4►
◄seed0►
◄use_pefalse►
◄s1_initial_mp0.40►
◄s1_denoise1.00►
◄s1_steps6►
◄s1_guidance_scale1.0►
◄s1_sigma_schedulefixed shift►
◄s1_shift_value6.0►
◄scale_factor2.5►
◄s2_denoise0.85►
◄s2_steps9►
◄s2_guidance_scale1.0►
◄s2_sigma_schedulefixed shift►
◄s2_shift_value5.0►
◄max_final_mp8.0►
◄upscale_methodbislerp►
◄guidance_rescale0.00►
◄vae_decode_fp32true►
◄s1_negative_prompt►
◄s2_negative_prompt►
CategoryEric ERNIE Image
Inputs (26)
| Name | Type | Default | Description |
|---|---|---|---|
| pipeline | ERNIE_PIPELINE | — | |
| prompt | STRING | Ignored when prompt_embeds connected. | |
| resolution | COMBO | 5:4 | Aspect ratio for both stages. |
| seed | INT | 00–18446744073709550000 | — |
| use_pe | BOOLEAN | false | — |
| s1_initial_mp | FLOAT | 0.400.2–1 | Stage 1 megapixels. 0.3-0.5 MP recommended. |
| s1_denoise | FLOAT | 1.000.05–1 | Stage 1 denoise strength - fraction of σ=[1.0→0] covered. 0.15 → stops at σ=0.85 (just lays in composition). |
| s1_steps | INT | 65–100 | Stage 1 steps. 6-8 for Turbo, 20-30 for SFT. |
| s1_guidance_scale | FLOAT | 1.00–20 | Stage 1 CFG. 1.0 for Turbo, 6-9 for SFT. |
| s1_sigma_schedule | COMBO | fixed shift | 'fixed shift' with s1_shift_value 6-8 keeps all steps structural. 'karras'/'beta' add detail-range emphasis on top of shift. |
| s1_shift_value | FLOAT | 6.00.1–16 | Stage 1 shift. 6-8 recommended. 1.0 = uniform. |
| scale_factor | FLOAT | 2.51–8 | Linear upscale factor. 2.5× on 0.4 MP ≈ 1 MP. Higher risks body horror. |
| s2_denoise | FLOAT | 0.850.1–1 | Noise level added to upscaled S1 latent before S2 denoise. 0.45 = 45% noise (light refinement, preserves S1 composition). 0.65+ = heavier re-generation with more new detail. 1.0 = full re-generation (ignores S1). |
| s2_steps | INT | 95–100 | Stage 2 detail steps. 6-10 for Turbo, 20+ for SFT. |
| s2_guidance_scale | FLOAT | 1.00–20 | Stage 2 CFG. 1.0 for Turbo, 2-3 for SFT. |
| s2_sigma_schedule | COMBO | fixed shift | 'fixed shift' with s2_shift_value for consistent noise schedule. 'beta' allocates more steps to fine-detail range. 'karras' emphasizes both extremes. |
| s2_shift_value | FLOAT | 5.00.1–8 | Stage 2 shift. 5.0 tested well with Turbo at 2.5× scale. |
| max_final_mp | FLOAT | 8.00.5–16 | Megapixel cap on final output. |
| upscale_method | COMBO | bislerp | bislerp (slerp-interpolation) preserves latent vector norms and angular relationships — sharper and more coherent than bicubic. |
| guidance_rescale | FLOAT | 0.000–1 | Rescale CFG output toward unit variance. Reduces saturation at high guidance_scale. 0.0 = off. 0.5-0.7 recommended for S1 cfg 6-9. Applied to both stages. |
| vae_decode_fp32 | BOOLEAN | true | Upcast VAE to float32 for decode. Prevents bf16/fp16 quantization of fine textures. Applies to both S1 preview and S2 final decode. |
| s1_negative_promptopt | STRING | Stage 1 negative prompt. | |
| s2_negative_promptopt | STRING | Stage 2 negative. Leave empty to reuse Stage 1. | |
| prompt_embedsopt | ERNIE_EMBEDS | Pre-computed positive embeddings. | |
| s1_negative_embedsopt | ERNIE_EMBEDS | Pre-computed Stage 1 negative embeddings. | |
| s2_negative_embedsopt | ERNIE_EMBEDS | Pre-computed Stage 2 negative embeddings. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| final_image | IMAGE | — |
| stage1_preview | IMAGE | — |