Laboratory for Image and Video Engineering (LIVE), The University of Texas at Austin
Advised by Prof. Alan C. Bovik
NeurIPS 2025 · November 2025
Saini, Gupta, Bovik. Rectified CFG++ for Flow Based Models. NeurIPS 2025 (arXiv:2510.07631)
The question
Why does guidance break on the best image generators?
Flux, SD3 and SD3.5 sample a rectified flow with a deterministic ODE. Classifier-free guidance (CFG) pushes their trajectories off the manifold, and nothing brings them back.
Why we need guidance at all
Without guidance, a flow model gives generic images that only loosely match the prompt.
Guidance should amplify the prompt: sharper detail, fuller color, better alignment.
CFG does that for diffusion models. On flow models it oversaturates.
Flux-dev, same prompt, same seed
No guidance
CFG, ω = 3.5
Ours, λ = 3.5
SD3.5, same prompt, same seed
No guidance
CFG, ω = 3.5
Ours, λ = 3
No guidance is dark and muddy. CFG oversaturates. Ours stays natural and faithful to the prompt.
Rectified-CFG++NeurIPS 2025
What CFG does to flow models
Standard CFG on random prompts. These are typical results.
Flux-dev with CFG, ω = 3 (paper Figure 6)
Plastic, cartoon look
Oversaturated
Flat, clipart look
Blown highlights
Washed out
SD3 with CFG, ω = 3 to 3.5
Oversaturated
Text garbled
Misspelled
Color blow-out
Color distortion
The same four failures every time: oversaturation, a plastic look, garbled text, flat texture.
Rectified-CFG++NeurIPS 2025
Why CFG works on diffusion but breaks on flows
Same extrapolation, different samplers
Sampler
Update
After CFG pushes the sample off the manifold
SDE (diffusion)
dx = [f(x,t) − g(t)2 ∇log pt(x)] dt + g(t) dw̄
fresh noise at every step partly pulls it back
ODE (rectified flow)
dx = vθ(x, t) dt
nothing pulls it back, and the error compounds
Discrete DDPM step: xt−1 = μθ(xt, t) + σt z with z ~ N(0, I). The σt z term is the fresh noise.
So we need guidance built for deterministic ODE samplers.
Rectified-CFG++NeurIPS 2025
What good guidance for a flow model should do
D1. Stay on the manifold, so errors cannot pile up.
D2. Guide strongly early (layout, color), gently late (text, texture).
D3. Come with provable bounds.
D4. Drop into any pretrained model, no retraining.
Method
D1
D2
D3
D4
CFG
no
no
no
yes
CFG-Zero*
no
no
no
yes
APG
partly
no
no
yes
Rectified-CFG++ (ours)
yes
yes
yes
yes
No existing method meets all four. APG is adaptive projected guidance.
We predict a midpoint on the manifold, evaluate guidance there, and add it as a bounded correction.
Rectified-CFG++NeurIPS 2025
Why CFG fails and how we fix it, in one picture
Ideal flow
The sample moves along the manifold.
CFG
Extrapolation steps off it, and the error stays.
Rectified-CFG++
Predict, then correct. The step lands back on the manifold.
Rectified-CFG++NeurIPS 2025
Our predictor-corrector stays on the manifold
1. Predict a half-step with vc alone
x̃mid = xt + (Δt/2) vct
2. Evaluate guidance at the midpoint
Δvmid = vcmid − vumid
3. Correct a bounded addition to vc
v̂ = vct + α(t)(vcmid − vumid)
α(t) = λmax(1−t)γ fades to 0 near the data, so fine detail survives.
The trajectory stays in a bounded tubular neighbourhood of Mt.
Rectified-CFG++NeurIPS 2025
Extrapolation versus interpolation
Where each rule puts the guided velocity
CFG extrapolates (ω > 1)
(1−ω)vu + ωvc with ω ∈ [3, 9]. ω = 3 means three times the distance from vu to vc, where the model never trained.
Ours interpolates (0 ≤ α ≤ 1)
vc + α(t) Δvmid with α ∈ [0, 1]. The base is always vc; the correction is bounded and decays to 0 near the data.
Rectified-CFG++NeurIPS 2025
The two algorithms side by side
Highlighted lines are what changes
Standard CFG
for t = T down to 1:
vc = vθ(xt, t, y)
vu = vθ(xt, t, ∅)
v̂ = vu + ω(vc − vu)extrapolation
xt−1 = xt + Δt v̂
2 network evaluations (NFE) per step
Rectified-CFG++
for t = T down to 1:
vc = vθ(xt, t, y)
x̃mid = xt + (Δt/2) vcpredictor
vcmid = vθ(x̃mid, t − Δt/2, y)
vumid = vθ(x̃mid, t − Δt/2, ∅)
v̂ = vc + α(t)(vcmid − vumid) interpolation
xt−1 = xt + Δt v̂
3 NFE per step
The base is always vc.
Guidance is evaluated at the predicted midpoint.
α(t) = λ(1−t)γ fades to 0 near the data.
Rectified-CFG++NeurIPS 2025
What we can prove
Four mild assumptions
A1
vθ(x,t,y) and vθ(x,t,∅) are L-Lipschitz in x
A2
Bounded guidance direction, ‖Δvtθ(x)‖ ≤ B
A3
α(t) is bounded and integrable
A4
Bounded conditional velocity, ‖vct‖ ≤ Vmax
Standard regularity conditions for a trained network.
Three bounds
Lemma 1 guidance stability
‖Δvθt−Δt/2 − Δvθt(xt)‖ ≤ L Vmax Δt
The midpoint moves the guidance direction by only O(Δt)
Proposition 1 single-step perturbation
‖x̂t−1 − x̃t−1‖ ≤ α(t)B Δt
Deviation from the conditional path vanishes near the data
Proposition 2 distributional deviation
DKL(p̂0 ‖ p0) ≤ Cs (B + L Vmax Δt) ∫α(τ) dτ
Total KL is set by the integrated guidance
No earlier guidance method for flow models has bounds like these.
Rectified-CFG++NeurIPS 2025
CFG versus ours, property by property
Lemma 2, manifold-faithful corrector
dist(x̂, Mt−Δt) ≤ C ε Δt
Only the training error ε moves us off the manifold.
Integrated guidance
4 to 16× lower KL divergence
Property
CFG
Ours
Per-step perturbation
(ω−1) B Δt constant, never vanishes
α(t) B Δt goes to 0 near the data
Manifold distance
Unbounded errors accumulate
O(ε Δt) training error only
Integrated guidance
ω−1 ≈ 2 to 8
λ/(γ+1) ≈ 0.5
KL(p̂0 ‖ p0)
O(ω−1)
O(0.5)
Late-step behavior
Same ω everywhere, destroys fine detail
α(t) goes to 0, preserves fine detail
Guidance evaluation
At the current point (possibly off the manifold)
At the predicted midpoint (on the manifold)
Rectified-CFG++NeurIPS 2025
Watching the artifacts grow, or not
Standard CFG, ω = 2.5. Each strip shows 7 denoising steps, noise on the left, final image on the right.
Rectified-CFG++, λ = 0.5, same prompt
CFG, another prompt. Drift starts at step 2 or 3.
Ours, same prompt. Every intermediate looks natural.
Rectified-CFG++NeurIPS 2025
A 2D toy problem shows the same drift
CFG overshoots and needs sharp late corrections. Ours converges smoothly.
t = 1.0t = 0.8t = 0.6t = 0.4t = 0.2t = 0.0
CFGRectified-CFG++
200 trajectories on a 2D mixed Gaussian. Top: CFG (blue) drifts off the support. Bottom: ours (green) stays on the manifold.
Rectified-CFG++NeurIPS 2025
Baseline, CFG and Rectified-CFG++ on three Flux prompts
Rows from the top: Flux with no guidance, CFG, Rectified-CFG++.
Cactus. CFG draws a cartoon with an emoji face.
Dog and moon gate. CFG loses the stone arch.
Parrot. CFG turns the bird psychedelic.
Ours keeps the realism and the structure in all three.
Rectified-CFG++NeurIPS 2025
The same on SD3 and Lumina-Next
SD3, CFG
SD3, ours
SD3, CFG
SD3, ours
Lumina, CFG
Lumina, ours
Lumina, CFG
Lumina, ours
A drop-in replacement across rectified flow models. No retraining, no architecture changes.
Rectified-CFG++NeurIPS 2025
Text survives because guidance fades at the end
CFG
Rectified-CFG++
Prompt: "A stop sign with 'ALL WAY' written below it."
α(t) is 0 in the last steps, so the conditional flow alone draws the strokes.
Rectified-CFG++NeurIPS 2025
Results on MS-COCO 10K
Model
Method
FID ↓
CLIP ↑
ImgRwd ↑
Pick ↑
HPS ↑
Lumina
CFG
26.93
0.261
0.547
21.03
0.253
Lumina
Ours
22.49
0.268
0.621
21.19
0.259
SD3
CFG
26.33
0.273
0.684
21.48
0.264
SD3
Ours
24.68
0.278
0.753
21.62
0.269
Flux
CFG
37.86
0.285
0.892
22.04
0.279
Flux
Ours
32.23
0.289
0.961
22.18
0.283
CFG against Rectified-CFG++ on three flow models. Bold is best.