ComfyUI-AutoGuidance

A ComfyUI custom node implementing AutoGuidance from the paper:

Guiding a Diffusion Model with a Bad Version of Itself (Karras et al., 2024)
https://arxiv.org/abs/2406.02507

Repository: https://github.com/xmarre/ComfyUI-AutoGuidance

Reddit Thread with more info on usage: https://www.reddit.com/r/StableDiffusion/comments/1r2a7qo/release_comfyuiautoguidance_guide_the_model_with/

What this does

Classic CFG steers generation by contrasting conditional and unconditional predictions.
AutoGuidance adds a second model path (“bad model”) and guides relative to that weaker reference.

In practice, this gives you another control axis for balancing:

• quality/faithfulness,
• collapse/overcooking risk,
• structure vs detail emphasis (via ramping).

Included nodes

This extension registers two nodes:

• AutoGuidance CFG Guider (good+bad) (AutoGuidanceCFGGuider)
  Produces a GUIDER for use with SamplerCustomAdvanced.
• AutoGuidance Detailer Hook (Impact Pack) (AutoGuidanceImpactDetailerHookProvider)
  Produces a DETAILER_HOOK for Impact Pack detailer workflows (including FaceDetailer).

Installation

Clone into your ComfyUI custom nodes directory and restart ComfyUI:

git clone https://github.com/xmarre/ComfyUI-AutoGuidance

No extra dependencies.

Basic wiring (SamplerCustomAdvanced)

1. Load two models:
   • good_model
   • bad_model
2. Build conditioning normally:
   • positive
   • negative
3. Add AutoGuidance CFG Guider (good+bad).
4. Connect its GUIDER output to SamplerCustomAdvanced guider input.

Impact Pack / FaceDetailer integration

Use AutoGuidance Detailer Hook (Impact Pack) when your detailer nodes accept a DETAILER_HOOK.

This injects AutoGuidance into detailer sampling passes without editing Impact Pack source files.

Important: dual-model mode must use truly distinct model instances

If you use:

• swap_mode = dual_models_2x_vram

then ensure ComfyUI does not dedupe the two model loads into one shared instance.

Recommended setup

Make a real file copy of your checkpoint (same bytes, different filename), for example:

• SDXL_base.safetensors
• SDXL_base_BADCOPY.safetensors

Then:

• Loader A (file 1) → good_model
• Loader B (file 2) → bad_model

If both loaders point to the exact same path, ComfyUI may share/collapse model state and dual-mode behavior/performance can be incorrect.

Parameters (AutoGuidance CFG Guider)

Required

• cfg
• w_autoguide (effect is effectively off at 1.0; stronger above 1.0)
• swap_mode
  • shared_safe_low_vram (safest/slowest)
  • shared_fast_extra_vram (faster shared swap, extra VRAM)
  • dual_models_2x_vram (fastest, highest VRAM, requires distinct instances)

Optional core controls

• ag_combine_mode
  • sequential_delta (default; current behavior: CFG result + AG delta)
  • multi_guidance_paper (paper-style multi-guide extrapolation using good-cond, good-uncond, bad-cond)
  • In multi_guidance_paper, ag_delta_mode, ag_max_ratio, ramping, and project/reject controls are not used.
• ag_delta_mode
  • bad_conditional (default, common starting point)
  • raw_delta
  • project_cfg
  • reject_cfg
• ag_max_ratio (caps AutoGuidance push relative to CFG update magnitude)
• ag_allow_negative
• ag_ramp_mode
  • flat
  • detail_late
  • compose_early
  • mid_peak
• ag_ramp_power
• ag_ramp_floor
• ag_post_cfg_mode
  • keep
  • apply_after
  • skip

Swap/debug controls

• safe_force_clean_swap
• uuid_only_noop
• debug_swap
• debug_metrics

Example setup (one working recipe)

This is one practical baseline, not a rule.

Models

• Good side:
  • Base checkpoint + more fully-trained/specialized stack (e.g., 40-epoch character LoRA + DMD2/LCM, etc.)
• Bad side options:
  • Base checkpoint + earlier/weaker checkpoint/LoRA (e.g., 10-epoch) with intentionally poor weighting
  • Base checkpoint + fewer adaptation modules
  • Base checkpoint only

Core idea: bad side should be meaningfully weaker/less specialized than good side.

Node settings example

• cfg: 1.1
• w_autoguide: 3.00
  • note: in multi_guidance_paper, values above 2.0 clamp to α=1.0 (pure AG share).
• swap_mode: dual_models_2x_vram
• ag_delta_mode: reject_cfg
• ag_max_ratio: 0.75
• ag_allow_negative: true
• ag_ramp_mode: compose_early
• ag_ramp_power: 2.0
• ag_ramp_floor: 0.00
• ag_post_cfg_mode: skip
• safe_force_clean_swap: true
• uuid_only_noop: false
• debug_swap: true
• debug_metrics: true

Practical tuning notes

• In non-paper modes, increase w_autoguide above 1.0 to strengthen effect.
• In multi_guidance_paper mode, weights are interpreted as:
  • α = clamp(w_autoguide - 1, 0, 1)
  • w_total = max(cfg - 1, 0)
  • w_cfg = (1 - α) * w_total
  • w_ag = α * w_total
  • CFG leg uses cfg_scale = 1 + w_cfg (including sampler CFG hooks).
  • final output = CFG(good, cfg_scale) + w_ag * (C - B) where C = good conditional and B = bad conditional.
• To set paper interpolation directly (total g, mix α):
  • cfg = g
  • w_autoguide = 1 + α
  • ag_combine_mode = multi_guidance_paper
• In multi_guidance_paper, ag_post_cfg_mode still works:
  • keep/skip: AG is mixed before post-CFG hooks.
  • apply_after: run post-CFG hooks on CFG-only output, then add AG delta.
• In multi_guidance_paper, ag_delta_mode, ag_max_ratio, ag_ramp_mode, ag_ramp_power, and ag_ramp_floor are currently ignored.
• compose_early tends to affect composition/structure earlier in denoise.
• Try detail_late for a more late-step/detail-leaning influence.

VRAM and speed

AutoGuidance adds extra forward work versus plain CFG.

• dual_models_2x_vram: fastest but highest VRAM and strict dual-instance requirement.
• Shared modes: lower VRAM, slower due to swapping.

Suggested A/B evaluation

At fixed seed/steps, compare:

• CFG-only vs CFG + AutoGuidance
• different ag_ramp_mode
• different ag_max_ratio caps
• different ag_delta_mode

Feedback wanted

Useful community feedback includes:

• what “bad model” definitions work best in real SD pipelines,
• parameter combos that outperform standard CFG or NAG,
• reproducible A/B examples with fixed seed + settings.

---

If this node helps, share workflows and side-by-side comparisons in issues/discussions.