NeurIPS 2025

Rectified-CFG++

Geometry-Aware Guidance for Rectified Flow Models

Shreshth Saini

Laboratory for Image & Video Engineering (LIVE) · The University of Texas at Austin

Advised by Prof. Alan C. Bovik

What starts here changes the world

Why Do We Need Guidance At All?

Without guidance, flow models generate plausible but generic images that loosely match the prompt

The problem with unconditional generation

Models trained with conditioning dropout learn both p(x|y) and p(x). Without guidance, outputs come from a mixture of conditional and unconditional — resulting in low prompt adherence, muted details, and bland compositions.

What guidance should do

Amplify the conditional signal — push generation toward what the prompt describes. Sharper details, more vivid colors, better text-image alignment. This is what CFG achieves in diffusion models (DDPM, SDXL).

What actually happens on flow models

CFG was designed for stochastic SDE samplers. Flow models use deterministic ODE. The extrapolation that works in SDEs becomes catastrophic in ODEs — artifacts, color blow-out, structural distortion.

Flux-dev: Same prompt, same seed

No guidance

CFG ω=3.5

Ours λ=3.5

No guidance: dark, muddy, poor prompt adherence. CFG: improved adherence but oversaturated, unnatural. Ours: clean, natural, faithful to prompt.

No guidance

CFG ω=3.5

Ours λ=3

SD3.5: Same pattern — CFG oversaturates, ours stays natural.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

CFG Artifacts on Flow Models: A Gallery of Failures

All images generated with standard CFG on flow models. These are not cherry-picked — this is what CFG does to every prompt.

Flux-dev with CFG ω=3 (from paper Figure 6)

Plastic / cartoon

Oversaturated

Flat / clipart

Blown highlights

Washed out

SD3 with CFG ω=3–3.5

Oversaturated

Text garbled

Misspelled

Color blowout

Color distortion

Common failure modes: (1) Oversaturation — colors pushed beyond natural range. (2) Cartoonification — photorealistic becomes plastic. (3) Text corruption — letters distort and merge. (4) Detail blowout — fine textures replaced by flat patches.

Not rare failures. CFG artifacts appear on every prompt across Flux, SD3, and SD3.5. The problem is fundamental: CFG extrapolation is incompatible with the deterministic ODE used by all modern flow models.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Root Cause: Why CFG Breaks on Flows but Works on Diffusion

The same extrapolation trick, but fundamentally different samplers

Obviously broken: CFG on SD3 & Flux

SD3: "CyberCore Cafe"

SD3: "Entropy and the Soul"

Flux: plastic/cartoon

Garbled text ("Cyberre Cidie"), misspellings ("Entφy amd"), plastic cartoonification — all from standard CFG.

SDE Sampler (Diffusion) vs ODE Sampler (Flows)

Diffusion SDEs: built-in safety net

Each step adds Gaussian noise — the renoising step. This noise accidentally provides error correction: even when CFG pushes off-manifold, the stochastic noise partially pulls the sample back toward the learned distribution.

dx = [f(x,t) − g(t)²·∇log p_t(x)]dt + g(t)·dw̄ ← stochastic noise term

Discrete (DDPM): x_t-1 = μ_θ(x_t, t) + σ_t·z, z ~ N(0, I). The σ_t·z term adds fresh noise at every step — pulling samples back toward the learned distribution.

Flow ODEs: no correction mechanism

Rectified flows use a purely deterministic ODE. Once CFG pushes the trajectory off the manifold at step t, there is no mechanism to return. The off-manifold point becomes the starting point for step t+1.

dx = v_θ(x, t)·dt ← purely deterministic, no noise term!

Compounding error

At each step, CFG evaluates guidance at an off-manifold point — making the velocity direction even more wrong. Errors accumulate monotonically across all N sampling steps. The more steps, the worse it gets.

This is why we need a fundamentally different approach — not a fix for CFG, but a new guidance paradigm designed specifically for deterministic ODE samplers.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

What Would an Ideal Guidance Method Look Like?

Desiderata for guidance on rectified flow models

D1: Stay on the manifold

The guided trajectory must remain in a bounded neighbourhood of the learned transport manifold M_t. No off-manifold drift, no error accumulation.

D2: Adaptive — strong early, gentle late

Early steps determine global structure (layout, color palette) — guidance should be strong. Late steps resolve fine details (text, textures) — guidance should vanish to prevent corruption.

D3: Provable guarantees

Not just empirically better — we want formal bounds on manifold distance and distributional deviation. No prior guidance method provides this.

D4: Drop-in replacement

Works with any pretrained flow model — Flux, SD3, SD3.5, Lumina — without retraining. Just swap the sampling algorithm.

Do existing methods satisfy these?

Method	D1	D2	D3	D4
CFG	✗	✗	✗	✓
CFG-Zero*	✗	✗	✗	✓
APG	~	✗	✗	✓
Rect-CFG++ (Ours)	✓	✓	✓	✓

Our key insight: Don't extrapolate the guidance direction from the current point. Instead, predict a midpoint on the manifold, evaluate guidance there, and apply it as a bounded additive correction to the conditional velocity.

Three ideas → three guarantees:

Predictor half-step (on-manifold evaluation) + Adaptive α(t) (vanishing guidance) + Bounded correction (interpolation not extrapolation)

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

The Geometric Insight: Why CFG Fails and How We Fix It

Three panels: (Left) Ideal flow on manifold. (Center) CFG extrapolates off-manifold. (Right) Rect-CFG++ stays on-manifold via predictor-corrector. Let's walk through each.

Why does CFG work in diffusion but not in flows?

In diffusion models (SDE samplers like DDPM), each step adds stochastic noise — the renoising step. This noise acts as implicit error correction: even if CFG pushes the trajectory off-manifold, the added noise partially brings it back. It's a built-in safety net.

In rectified flow models (ODE solvers like Flux), sampling is purely deterministic — there is no renoising step. Once CFG pushes the trajectory off the manifold, there is no mechanism to return. The error at step t becomes the starting point for step t-1, and the next CFG extrapolation makes it worse. Errors accumulate monotonically across all sampling steps.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Geometric Insight (1/3): The Ideal Flow on the Manifold

What you're seeing

Three manifolds: M_t (current time, blue/gold band), M_t-1 (next step), M₀ (data, orange curve at bottom). The red dots are the ideal positions at each timestep.

The ideal trajectory

The ideal sample flows along the manifold surface from M_t → M_t-1 → M₀. At each step, the velocity lies in the tangent space of the manifold — never leaving it.

Three flows shown

Blue = Flux conditional (v^c), Red = CFG (extrapolated), Green = Rectified-CFG++ (ours). Notice how green and blue stay near the manifold while red diverges.

Key question

How do we guide the generation (improve text adherence) while staying on the manifold? Let's see what goes wrong first...

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Geometric Insight (2/3): CFG Extrapolates Off-Manifold

What CFG does

Starting from the black dot on M_t, CFG computes the guided velocity v̂_ω at the current point. With ω>1, this velocity extrapolates past the conditional velocity v^c.

v̂_ω = v^u + ω(v^c − v^u) ω > 1 → overshoots!

The tangent space violation

The guided velocity v̂_ω does not belong to the tangent space T_{x_t}M_t. The resulting Euler step lands the sample off M_t-1. In a deterministic ODE, there is no mechanism to return.

The compounding effect

At the next step, CFG evaluates guidance at the off-manifold point — making the direction even worse. Errors accumulate at every step. By the final image: oversaturation, color blow-out, structural distortion.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Geometric Insight (3/3): Our Predictor-Corrector Stays On-Manifold

Step 1: Predict midpoint with v^c only

Take a half-step using only the conditional velocity: x̃_mid = x_t + (Δt/2)·v^c_t. Since v^c lies in the tangent space, x̃_mid lands on or near M_t-Δt/2.

Step 2: Evaluate guidance at the predicted midpoint

Compute v^c_mid and v^u_mid at x̃_mid. The guidance direction Δv_mid = v^c_mid − v^u_mid is evaluated on the manifold, reflecting the geometry of the target manifold M_t-Δt.

Step 3: Apply bounded interpolative correction

v̂ = v^c_t + α(t)(v^c_mid − v^u_mid)

α(t) = λ_max(1-t)^γ. Base velocity is always v^c (on manifold). Correction is additive and bounded. α(t)→0 near data → fine details preserved.

Result: trajectory stays in bounded tubular neighbourhood of M_t

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

The Core Distinction: Extrapolation vs. Interpolation

This single difference explains why CFG fails and Rect-CFG++ works

CFG: Extrapolation (ω > 1)

CFG moves PAST the conditional direction.
ω=3 means 3× the distance from v^u to v^c. Overshoots into unknown territory.

Ours: Interpolation (0 ≤ α ≤ 1)

Ours stays BETWEEN v^c and v^c+Δv.
Base is always v^c (on manifold). Correction α(t)·Δv_mid is bounded and decays to 0 near data.

Mathematical contrast: CFG: (1−ω)v^u + ωv^c with ω∈[3,9] → extrapolation far past v^c | Ours: v^c + α(t)Δv_mid with α∈[0,1] → bounded additive correction to v^c

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Algorithm: CFG vs. Rectified-CFG++ (Side-by-Side)

Color highlights show exactly what changes — green = our additions

Standard CFG Sampling

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 NFE/step. ω>1 pushes off manifold. No correction.

Rectified-CFG++ (Ours)

for t = T down to 1:
  vc = vθ(xt, t, y)
  x̃mid = xt + (Δt/2)·vc ← predictor
  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/step. α(t)=λ(1-t)^γ decays → on-manifold. Bounded.

Key differences: (1) Base velocity is always v^c, not extrapolated. (2) Guidance evaluated at predicted midpoint on manifold, not current point. (3) Adaptive α(t) → 0 near data.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Guidance Methods Compared: Equations & Properties

Four approaches to the same problem — only ours avoids extrapolation entirely

Standard CFG (Ho & Salimans, 2022)

v̂ = (1 − ω)·v^u(x_t) + ω·v^c(x_t)

ω > 1 → extrapolation past v^c. Designed for SDEs with renoising. On ODEs: off-manifold drift, oversaturation, error accumulation. Constant guidance at all timesteps.

CFG-Zero* (Wang et al., 2024)

v̂ = (1 − ω)·s_t^*·v^u(x_t) + ω·v^c(x_t)

Adds an optimal rescaling factor s_t* to align unconditional/conditional norms. Reduces initial drift but still extrapolates (ω > 1). No late-step adaptation — fine details still corrupted.

APG (Sadat et al., 2024)

v̂ = v^c(x_t) + Δv_t^(η,r,β)

Projects guidance direction onto perpendicular subspace to reduce parallel component. Partial mitigation but guidance still evaluated at current point (possibly off-manifold). No decay schedule, no formal bounds.

Rectified-CFG++ (Ours)

v̂ = v^c(x_t) + α(t)·(v^c_mid − v^u_mid)

Three key differences: (1) Base = v^c always (on-manifold). (2) Guidance at predicted midpoint x̃_mid (on-manifold). (3) α(t) = λ(1-t)^γ → 0 near data. Bounded, interpolative, provable.

Property	CFG	CFG-Zero*	APG	Ours
Guidance type	Extrapolation	Extrapolation	Projection	Interpolation
Evaluated at	Current x_t	Current x_t	Current x_t	Predicted x̃_mid
Temporal schedule	None (constant)	None	None	α(t) → 0
Formal guarantees	✗	✗	✗	✓ (3 props)
NFE per step	2	2	2	3

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Theoretical Guarantees — First Provable Bounds for Flow Guidance

Four mild assumptions → three rigorous guarantees. No prior guidance method has these.

Assumptions

(A1) v_θ(x,t,y) and v_θ(x,t,∅) are L-Lipschitz in x

(A2) Guidance direction bounded: ‖Δv_t^θ(x)‖ ≤ B

(A3) Schedule α(t) is bounded and integrable

(A4) Conditional velocity bounded: ‖v^c_t‖ ≤ V_max

All standard regularity conditions. Lipschitz and boundedness are satisfied by any well-trained neural network. No exotic assumptions needed.

Lemma 1: Guidance Stability

‖Δv^θ_t-Δt/2 − Δv^θ_t(x_t)‖ ≤ L · V_max · Δt

Guidance direction at the predicted midpoint differs from guidance at current point by only O(Δt). The predictor doesn't introduce large errors.

Proposition 1: Bounded Single-Step Perturbation

‖x̂_t-1 − x̃_t-1‖ ≤ α(t) · B · Δt

Deviation from the pure conditional path is controlled by α(t) at each step. Since α(t) → 0 as t → 0, perturbations vanish near the data manifold.

Proposition 2: Distributional Deviation

D_KL(p̂₀‖p₀) ≤ C_s·(B + LV_maxΔt)·∫α(τ)dτ

Total KL divergence controlled by integrated guidance ∫α(τ)dτ — the single quantity that determines output quality.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Why These Guarantees Matter: CFG vs. Ours

Lemma 2: Manifold-Faithful Corrector

dist(x̂, M_t-Δt) ≤ C·ε·Δt

Distance to manifold bounded by training error ε only. The corrector displacement is tangent to M_t-Δt/2 — guidance doesn't push off-manifold.

Integrated Guidance: The Key Quantity

Head-to-Head: Every Property

Property	CFG	Ours
Per-step perturbation	(ω−1)·B·Δt constant, never vanishes	α(t)·B·Δt → 0 near data
Manifold distance	Unbounded errors accumulate	O(ε·Δt) training error only
Integrated guidance	ω−1 ≈ 2–8	λ/(γ+1) ≈ 0.5
KL(p̂₀ ‖ p₀)	O(ω−1)	O(0.5)
Late-step behavior	Same ω everywhere → destroys fine details	α(t) → 0 → preserves fine details
Guidance evaluation	At current point (possibly off-manifold)	At predicted midpoint (on manifold)

Bottom line: CFG has no convergence guarantee for flow models. Ours proves that the output distribution deviates from the true conditional distribution by a controllable, bounded amount — proportional to ∫α(τ)dτ which we set to ≈0.5.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Intermediate Denoising: Watching Artifacts Grow vs. Stay Clean

Each strip shows 7 denoising steps from noise (left) to final image (right)

Standard CFG (ω=2.5) — artifacts compound

Rectified-CFG++ (λ=0.5) — clean throughout

CFG: Off-manifold drift begins at early steps (step 2-3). By mid-trajectory, color artifacts are baked in. Late steps cannot correct — each step amplifies the error. The final image has oversaturated colors and unnatural contrast.

Ours: Predictor step keeps each intermediate on M_t. α(t) schedule applies strong guidance early (global structure) and vanishes late (fine details). Result: every intermediate looks natural. No error accumulation.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

2D Toy: CFG Drifts Off-Support, Ours Stays On-Manifold

200 trajectories on 2D mixed Gaussian. Top: CFG (blue) drifts off support. Bottom: Ours (green) smooth on-manifold.

CFG (top)

Trajectories initially overshoot, leaving the learned transport manifold. Sharp late-stage corrections pull samples back — but damage is done. Final distribution is noisy and off-center.

Rectified-CFG++ (bottom)

Smooth convergence throughout. Predictor anchors each step on the flow field. Corrector applies bounded guidance. Samples arrive at target with tight concentration — no drift, no sharp corrections.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Flux: Baseline vs. CFG vs. Rectified-CFG++ (3 Prompts)

Top: Flux baseline (no guidance). Middle: CFG — oversaturated, cartoonish, structural distortion. Bottom: Rect-CFG++ — sharp details, faithful colors, correct structure.

Cactus: CFG turns photorealistic cactus into cartoon with emoji-like face. Ours preserves desert realism.

Dog + moon gate: CFG loses the stone arch, replaces with painted mountains. Ours preserves the gate structure and moon.

Parrot: CFG turns the bird into a psychedelic abstraction. Ours preserves natural plumage and forest setting.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Flux: Baseline vs. CFG vs. Rectified-CFG++ (Same Prompt, Same Seed)

Same model (Flux-dev), same prompt, same seed — three guidance strategies. Left: no guidance (baseline). Center: CFG (ω=3.5). Right: Rectified-CFG++ (λ=0.5). Ours: sharper details, faithful colors, no rainbow artifacts.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Diverse Prompt Gallery: Rectified-CFG++ Handles Everything

All images generated with Rectified-CFG++ on Flux-dev. No cherry-picking — random prompts from MS-COCO and Pick-a-Pic.

Photorealistic scenes

Landscapes, portraits, food, architecture — natural lighting, correct shadows, no oversaturation. The on-manifold property preserves the photorealistic distribution Flux was trained on.

Artistic & stylized

Oil paintings, digital art, fantasy scenes — style transfer works correctly because α(t) allows strong guidance early (global style) while preserving fine brush strokes late.

Text & compositional

Signage, book titles, multi-object scenes — the hardest category for any guidance method. α(t)→0 at late steps preserves pixel-precise text rendering and spatial relationships.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

SD3.5: 4-Way Guidance Comparison (Same Prompt, Same Seed)

CFG ω=3.5

CFG-Zero*

APG

Ours λ=6

CFG

Ours

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Flux-dev: CFG vs. Ours (8 Comparisons)

CFG

Ours

CFG

Ours

CFG

Ours

CFG

Ours

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

More Visual Comparisons: SD3 & Lumina-Next

SD3 CFG

SD3 Ours

SD3 CFG

SD3 Ours

Lumina CFG

Lumina Ours

Lumina CFG

Lumina Ours

Drop-in replacement across all rectified flow models — no retraining, no arch changes.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Text Rendering: The Stop Sign Test

Text-heavy prompts are especially sensitive to off-manifold drift at late denoising steps

CFG

Rectified-CFG++

Prompt: "A stop sign with 'ALL WAY' written below it."

CFG: "STOP" text distorts — letter shapes warp, "ALL WAY" becomes unreadable. Off-manifold drift corrupts high-frequency text strokes in the final denoising steps.

Ours: Crisp, photorealistic stop sign. α(t)→0 near t=0 ensures pure conditional flow during fine detail resolution — text strokes remain pixel-precise.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Text Rendering: Neon Street Sign

CFG

Rectified-CFG++

Prompt: "A neon street sign that says 'CyberCore Cafe', glowing in magenta and blue."

CFG garbles the neon letters — "Cyberre Cidie Cafe" instead of "CyberCore Cafe". Glow halos and letter boundaries bleed together. Ours renders each letter distinctly with clean neon glow separation.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Text Rendering: Headlines & Inscriptions

CFG

Ours

CFG

Ours

Prompt: "A crow detective reading a paper titled 'Feathered Conspiracies', headline in bold noir font."

Prompt: "A magical sword embedded in stone, with the name 'SOLARFANG' etched along its blade in glowing runes."

CFG: "Feathered Conspiracies" → "Feathhrad Conspiracies". "SOLARFANG" → garbled runes. Ours: Both headlines render correctly with proper typography — even on complex surfaces like aged paper and glowing stone.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Text Rendering: Presentations & Artistic Text

CFG

Ours

CFG

Ours

Prompt: "A fox giving a TED talk, slide behind reads 'Entropy and the Soul'."

Prompt: "A burning scroll with the title 'Soulbound by Nightfall' in ornate calligraphy."

CFG: "Entropy and the Soul" → "Entφy amd the Soul". Title on scroll barely readable. Ours: Clean title text on both the presentation screen and the burning scroll — fine details preserved by zero guidance at final steps.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Text Rendering: Carved & Printed Text

CFG

Ours

CFG

Ours

Prompt: "A desert gate carved in sandstone reading 'CITY OF WHISPERS'."

Prompt: "A Japanese fish stamp print with 'WAVES OF JUSTICE' in red ink."

Pattern: Across all text types — neon, carved stone, ink stamps, calligraphy, digital displays — Rectified-CFG++ consistently preserves legibility. The adaptive schedule α(t)→0 is the key mechanism: no guidance perturbation during fine-detail resolution.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Text Legibility: Full Gallery (SD3.5)

Each column: CFG left → Rectified-CFG++ right

CFG

Ours

CFG

Ours

CFG

Ours

CFG

Ours

CFG

Ours

CFG

Ours

CFG

Ours

CFG

Ours

Consistent text legibility across all prompts — the adaptive schedule α(t) is the key mechanism.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Quantitative: MS-COCO 10K — Consistent Improvements

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

Guidance (SD3.5)	FID	ImgRwd	CLIP	HPSv2
CFG	67.71	1.053	0.352	0.294
CFG-Zero*	68.39	0.995	0.346	0.288
APG	67.23	1.075	0.351	0.294
Rect-CFG++	67.15	1.085	0.351	0.296

Best or tied on every metric, every model.

-5.6FID on Flux

-4.4FID on Lumina

4Models tested

0Extra compute

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

User Study: 30 Experts × 32 Prompts × 4 Models = 15,360 Responses

4-way forced choice: CFG (dark blue), APG (medium blue), CFG-Zero* (light blue), Rectified-CFG++ (green). 4 criteria × 4 models. Green bar is tallest in every single comparison.

Overall: 34-36% preference share across all models (random chance = 25%). p < 0.001 vs second-best (APG).

Text legibility: Highest improvement dimension — 28-32% vs 22-25% for alternatives. α(t) schedule is critical.

Protocol: Fleiss' κ = 0.61 (substantial agreement). Images randomized. Experts with CV/generative AI knowledge.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Ablation 1: Guidance Scale — CFG Crashes, Ours Stays Stable

X-axis: guidance scale (λ for ours | ω for CFG). Y-axis: FID↓, CLIP↑, ImageReward↑, Aesthetic↑. Blue = CFG, Green = Ours.

CFG (blue curves)

FID explodes: 77 → 148 at ω=10. ImageReward collapses: 1.0 → -1.5. CLIP drops 30%. Aesthetic drops 35%. Catastrophic failure at high guidance.

Rectified-CFG++ (green curves)

FID: 75 → 75 (flat!). ImageReward: 1.0 → 0.95. CLIP stable. Aesthetic stable. Graceful degradation — never crashes.

Why this matters

CFG requires careful per-model tuning of ω to avoid collapse. Rect-CFG++ is robust across a 50× range of λ — no hyperparameter sensitivity. One setting works everywhere.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Ablation 2: Sampling Steps — Better Quality with Fewer Steps

X-axis: sampling steps. Blue = CFG, Green = Ours. Our method converges faster across all metrics.

Ours at 5 steps vs CFG at 5 steps

FID: 71 vs 178 (2.5× better). ImageReward: 1.0 vs -1.5. At ultra-low NFE, CFG produces garbage; ours produces usable images.

Ours at 15 steps ≈ CFG at 28 steps

Same FID, same CLIP, same ImageReward — but ~2× faster inference. This is because our on-manifold trajectories converge more efficiently.

Component ablation (SD3.5)

Config	FID	CLIP	HPSv2
Unconditional only	91.12	0.144	0.187
w/o Predictor	73.70	0.341	0.297
w/o Corrector	74.65	0.341	0.298
Full	72.97	0.345	0.300

Both predictor and corrector essential.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin

What starts here changes the world

Key Takeaways

1

First on-manifold guidance for rectified flow models. Predictor-corrector with adaptive decay. Provably bounded manifold distance.

2

4-16× tighter distributional bounds than CFG. Three propositions with formal proofs. Integrated guidance λ/(γ+1) vs ω-1.

3

Universal drop-in, zero overhead. Flux, SD3, SD3.5, Lumina — all improve. No retraining. 15 steps ≈ CFG at 28.

4

43.5% preference, 15K expert responses. Best on detail, color, text, overall across all models. p < 0.001.

Next: Rect-CFG++ for video generation (Sora-class) · Distilled flow models · Integration with LumaFlux for guided HDR generation.

Rectified-CFG++ (NeurIPS '25)LIVE Lab, UT Austin