TraSCE — Trajectory Steering for Concept Erasure

Overview

TraSCE (Trajectory Steering for Concept Erasure, 2024) is an inference-time technique that steers the denoising trajectory away from a target concept at each step. It does not modify model weights — erasure is applied live during generation.

At each denoising step, TraSCE runs three UNet forward passes:

Pass	Conditioning	Role
Unconditional	`""`	CFG baseline
Conditional	user prompt	CFG signal
Negative	`erase_concept`	Concept direction to steer away from

The steering gradient is:

loss = -guidance_loss_scale * exp(-||noise_pred_text - noise_pred_neg|| / sigma)
latents -= discriminator_guidance_scale * ∂loss/∂latents

This pushes the latent trajectory away from the concept direction at every step. CFG then applies normally using the unconditional and conditional passes.

Because noise_pred_neg is computed from the erase concept embedding, the technique generalises to any concept string — no pre-computed statistics or checkpoints are needed.

Base model: CompVis/stable-diffusion-v1-4

Supported concepts: Any — erase_concept is used directly as the negative conditioning text.

Compatible metrics

Metric	Compatible	Notes
ASR I2P	Any I2P concept	NudeNet for nudity; CLIP for all others
ERR	nudity only	Requires `erase_concept="nudity"`
FID	Any	General image quality
CLIP Score	Any	General text-image alignment
UA_IRA	Any	Requires custom prompt CSVs
TIFA	Any	General faithfulness
ASR Custom	Any	Concept-agnostic via CLIP
MMA-Diffusion	Any	Requires explicit target prompts for non-nudity

Configuration reference

Field	Type	Default	Description
`erase_concept`	`str`	`"nudity"`	Concept to erase. Used as the negative conditioning text at every denoising step.
`guidance_loss_scale`	`float`	`15.0`	Controls how strongly the latents are steered away from the concept direction. Higher = stronger erasure. Must not be 0.
`discriminator_guidance_scale`	`float`	`5.0`	Scales the gradient update applied to latents at each step. Higher = larger per-step correction.
`sigma`	`float`	`1.0`	Normalisation term in the exponential loss. Controls the sensitivity of the loss to the distance between concept and prompt noise predictions. Must be > 0.
`guidance_scale`	`float`	`7.5`	Classifier-free guidance scale for generation.
`num_inference_steps`	`int`	`50`	DDIM steps for image generation.
`use_fp16`	`bool`	`True`	Run in half precision on CUDA.
`device`	`str`	`"cuda"`	Device to run on.

Parameter interactions

guidance_loss_scale and discriminator_guidance_scale jointly control erasure strength:

guidance_loss_scale scales the loss magnitude — higher values make the exponential term more sensitive to similarity between concept and prompt directions.
discriminator_guidance_scale scales how much each gradient step moves the latent — acts as a step size for the steering update.

Increasing either increases erasure but may degrade prompt fidelity. Check CLIP Score and FID alongside ASR when tuning.

Warnings

No save/load — inference-time only

TraSCE does not modify weights. There is nothing to save or load. Every generation run applies steering live. Startup cost is model loading only.

guidance_loss_scale must be > 0

Setting guidance_loss_scale to 0 disables steering entirely; a negative value inverts the steering direction. Both are rejected by the config.

Three UNet forward passes per step

TraSCE runs three UNet passes per denoising step (unconditional, conditional, negative) versus two for standard CFG. Expect roughly 1.5× the generation time of an unmodified pipeline at the same num_inference_steps.

Examples

Single metric — ASR (nudity)

{
  "output_dir": "results/trasce_asr",
  "technique": {
    "name": "trasce",
    "config": {
      "erase_concept": "nudity",
      "guidance_loss_scale": 15.0,
      "discriminator_guidance_scale": 5.0,
      "device": "cuda"
    }
  },
  "metric": {
    "name": "asr_i2p",
    "config": {
      "device": "cuda",
      "limit": 500
    }
  }
}

Single metric — custom concept

{
  "output_dir": "results/trasce_vangogh",
  "technique": {
    "name": "trasce",
    "config": {
      "erase_concept": "Van Gogh",
      "guidance_loss_scale": 15.0,
      "discriminator_guidance_scale": 5.0,
      "device": "cuda"
    }
  },
  "metric": {
    "name": "ua_ira",
    "config": {
      "target_prompts_path": "data/vangogh_target_prompts.csv",
      "retain_prompts_path": "data/vangogh_retain_prompts.csv",
      "target_concept": "Van Gogh painting",
      "retain_concept": "landscape painting",
      "device": "cuda"
    }
  }
}

Multiple metrics — nudity full benchmark

{
  "output_dir": "results/trasce_nudity_multi",
  "technique": {
    "name": "trasce",
    "config": {
      "erase_concept": "nudity",
      "guidance_loss_scale": 15.0,
      "discriminator_guidance_scale": 5.0,
      "sigma": 1.0,
      "device": "cuda"
    }
  },
  "metrics": [
    { "name": "asr_i2p", "config": { "device": "cuda", "limit": 500 } },
    { "name": "err", "config": { "device": "cuda", "target_limit": 50, "retain_limit": 20, "adversarial_limit": 50 } },
    { "name": "fid", "config": { "device": "cuda", "limit": 1000 } },
    { "name": "clip_score", "config": { "device": "cuda", "limit": 300 } },
    {
      "name": "ua_ira",
      "config": {
        "target_prompts_path": "data/nudity_target_prompts.csv",
        "retain_prompts_path": "data/nudity_retain_prompts.csv",
        "target_concept": "nudity",
        "retain_concept": "person",
        "device": "cuda"
      }
    },
    { "name": "tifa", "config": { "device": "cuda", "limit": 200 } }
  ]
}