Mode 2 · IMG2IMG — the same sampler, a different initial condition
Mode 2 is particularly useful as a learning example: most of the production system stays the same while the latent source and denoise change. A mode switch therefore does not require a second workflow — it can reconfigure the contract of an existing module.
Mode 1 and Mode 2 differ primarily in starting latent and denoise
| Parameter | Mode 1 · TXT+CNET2IMG | Mode 2 · IMG+CNET2IMG |
|---|---|---|
| Latent source | Empty latent | VAE-encoded input image |
| Denoise | 1.0 | User control 608; stored 0.3 |
| Prompt / ControlNet / model | Shared system | Shared system |
| Sampler/output contract | Same module role | Same module role |
IMG2IMG = generation with memory of the source image
In txt2img, the latent starts as an empty/noise canvas. In img2img, the source IMAGE is first encoded by the VAE into LATENT and the sampler modifies an existing representation.
Denoise controls how much freedom the process has to change that representation. Lower denoise generally increases the influence of the initial latent; this is a useful mental model, not an absolute guarantee of geometry preservation.
One switch changes several downstream decisions
Node 541 does more than label the mode. It affects the latent route and denoise logic. This is a classic shared authoritative control: one production decision propagates into several technical branches.
Exercise: explain Mode 2 without Hansen node names
If you can explain Mode 2 as “input image → encode → latent → partial denoise → decode,” you understand the transferable principle.