[Vamsi-2581]

Summary

• Add upscaler=on|off to --conv-direct for the standalone ESRGAN upscaler.
• Preserve --diffusion-conv-direct behavior when --conv-direct is omitted, and apply the setting in both CLI and server paths.
• Update CLI help, documentation, and deprecated-option guidance.

Closes #2163.

Review follow-up

The latest commits address the maintainer's concern about examples calling core/ internals: the example now parses the assignment locally. Please review the updated parser and behavior, then re-review this PR.

Verification

Not run in this cloud workflow. The repository contribution guide says not to include test code in PRs. Human review of the updated implementation and the remaining CI check is pending.

AI assistance: OpenAI Codex assisted with implementation and review. Human review is pending.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[vmobilis]

@Vamsi-2581, hello, maybe the parser could use same logic as for other conv-direct options, that is:

1. --conv-direct is checked for correctness of syntax with others in common.cpp#L750 and put into conv_direct_mode;
2. the conv_direct_mode is parsed into conv_direct_assignment in diffusion_engine.cpp#L918;
3. each module gets its own conv-direct flag in diffusion_engine.cpp#L1227.

[Vamsi-2581]

Thanks, the shared assignment path makes sense for runners initialized by StableDiffusionGGML. The ESRGAN upscaler here is created separately through new_upscaler_ctx, outside diffusion_engine.cpp, and the earlier review specifically asked examples not to call core internals. How would you prefer the standalone context to receive the parsed upscaler setting: should this be exposed through a public parameter/helper, or do you have another intended route? I’m holding this follow-up until I align with that boundary.

[vmobilis]

@Vamsi-2581, since the upscaler is running after the generation, maybe just set inside diffusion_engine.cpp the ctx_params.upscaler_conv_direct flag, and it can later be checked by the upscaler module.

And for upscale mode, skip all generation runners and parse conv-direct option only to configure ctx_params.upscaler_conv_direct.