SPEED Node Port Accelerates Krea-2
aoleg77 · reddit · 2026-07-13
This post introduces a project porting the SPEED node to Forge Neo and SwarmUI, aiming to make Krea-2 generation faster and more stable.
Core Approach
- Instead of generating at full resolution from the start, it runs the initial steps at a lower resolution before switching back to full resolution for the remaining steps.
- This multi-stage strategy delivers an overall speedup of roughly 1.25x–1.33x.
- Direct high-resolution generation often causes issues like "duplicated objects or extra limbs." Running partial steps at a lower resolution significantly mitigates this.
Difference from hires.fix
- hires.fix generates a complete low-resolution image first, then performs VAE decoding, upscales, re-encodes, and adds more steps at high resolution.
- SPEED is more direct: it transitions straight from low-res to high-res steps, with the latent decoded only once at the very end.
Additional Info
- The author isn't the original creator of this concept; they simply ported the existing SPEED node to Forge Neo / SwarmUI and tested it with Krea-2.
- The post notes that more complex multi-stage configurations are possible, but the author found the payoff minimal and noticeably slower.
More from coding & agent
- An MCP server signs every AI agent tool call into a verifiable Merkle chain — Funky_Chicken_22 · 2026-07-22
- Claude Code skill uses 10 Markdown rules to make outputs ADHD-friendly — alex_verem · 2026-07-22
- A Firecracker-based platform says it can host 6,000 AI agents on one 256 GB server — maritime_sh · 2026-07-22
- A better path to agent autonomy is running waves, finding friction, and iterating — JnBrymn · 2026-07-22
- AI agent designers map the visual and tonal cues behind companionship products — Unlikely-Platform-47 · 2026-07-22
- Coding agents are heading toward an AI-writes, AI-reviews, human-approves workflow — aftahi_ai · 2026-07-22