feat: add flux-2-max and gpt-image-1.5 models

[Kamilbenkirane]

Adds two new image generation models:

• flux-2-max: BFL FLUX.2 Max model for image generation
• gpt-image-1.5: OpenAI GPT Image 1.5 model

Related to:

• https://withceleste.ai/models/flux-2-max
• https://withceleste.ai/models/gpt-image-1.5

[claude]

PR Review: Add flux-2-max and gpt-image-1.5 models

Overview

This PR adds two new image generation models to the Celeste Python SDK. The changes are well-structured and follow existing patterns in the codebase.

---

✅ Code Quality & Best Practices

Strengths:

• Consistent with existing patterns: Both model definitions follow the established structure seen in other models (flux-2-pro, dall-e-3, gpt-image-1)
• Proper imports: All necessary constraint types are correctly imported
• Good documentation: Inline comments explain model-specific behavior (e.g., "flux-2-max always upsamples prompts")
• Clean organization: Models are added to the appropriate provider files

Minor Observations:

1. Model ordering: flux-2-max is added before flux-2-pro in the MODELS list. Consider if there's a preferred ordering convention (e.g., chronological, alphabetical, or by capability tier). Looking at the OpenAI models, they appear roughly chronological.

2. Display name formatting: The naming is inconsistent across the codebase:

   • BFL models use bracket notation: FLUX.2 [max], FLUX.2 [pro]
   • OpenAI models use plain text: GPT Image 1.5, DALL-E 3

   This is fine as it reflects provider branding, but worth noting.

---

🐛 Potential Issues

flux-2-max (packages/capabilities/image-generation/src/celeste_image_generation/providers/bfl/models.py:9-36):

1. ⚠️ Duplicate constraint definitions: The dimension constraints, presets, seed, safety_tolerance, and output_format are identical to flux-2-pro (lines 37-63). This creates maintenance burden - if these constraints change, they need updating in multiple places.

   Recommendation: Consider extracting common constraints into a shared constant:

   _FLUX_2_COMMON_DIMENSIONS = Dimensions(
       min_pixels=64 * 64,
       max_pixels=2048 * 2048,
       min_aspect_ratio=9 / 21,
       max_aspect_ratio=21 / 9,
       presets={...}
   )

2. ℹ️ Comment about grounding search: Line 31 mentions "Includes grounding search for real-time information integration" but this isn't reflected in the parameter constraints. This appears to be a model capability that's always enabled (like prompt upsampling), which is fine, but the comment could be clearer about whether this requires any API configuration.

gpt-image-1.5 (packages/capabilities/image-generation/src/celeste_image_generation/providers/openai/models.py:60-72):

1. ✅ streaming=False: Good catch setting this explicitly. Consistent with the model's capabilities.

2. ⚠️ Parameter constraints match gpt-image-1: The constraints are identical to gpt-image-1 (lines 34-42) except for the streaming capability. This raises a question: what differentiates v1.5 from v1? Is it quality, speed, or features not exposed in the API parameters?

3. ❓ Missing PARTIAL_IMAGES parameter: Both gpt-image-1 and gpt-image-1-mini support PARTIAL_IMAGES: Range(min=0, max=3) (lines 35, 50). Should gpt-image-1.5 support this as well, or is its non-streaming nature the reason for omitting it?

---

🔒 Security Concerns

✅ No security issues identified

• Using established constraint types prevents injection attacks
• No dynamic code execution or unsafe operations
• Parameters are properly validated through the constraint system

---

⚡ Performance Considerations

✅ No performance concerns

• Model definitions are loaded once at import time
• No computational overhead in the additions
• Constraint validation is efficient

---

🧪 Test Coverage

⚠️ Missing test coverage for new models

The integration test at packages/capabilities/image-generation/tests/integration_tests/test_image_generation/test_generate.py includes test cases for:

• flux-2-pro (BFL)
• dall-e-2 (OpenAI)
• imagen-4.0-fast-generate-001 (Google)
• seedream-4-0-250828 (BytePlus)

Recommendations:

1. Add integration test cases for the new models:

   (Provider.BFL, "flux-2-max", {"aspect_ratio": "1024x1024"}),
   (Provider.OPENAI, "gpt-image-1.5", {"aspect_ratio": "1024x1024", "quality": "medium"}),

2. The models should be tested to verify:

   • Model registration works correctly
   • Parameters are properly constrained
   • Image generation completes successfully
   • Output format matches expectations

---

📋 Summary

Verdict: ✅ Approve with minor recommendations

The code is production-ready and follows established patterns. The main suggestions are:

1. High Priority: Add integration tests for both new models
2. Medium Priority: Consider DRYing up duplicate constraints in BFL models
3. Low Priority: Clarify the differences between gpt-image-1.5 and gpt-image-1 (especially around PARTIAL_IMAGES support)

The additions are safe to merge, though test coverage would strengthen confidence in the implementation.

---

📚 Additional Context

Based on the PR links:

• https://withceleste.ai/models/flux-2-max
• https://withceleste.ai/models/gpt-image-1.5

It would be helpful to verify the parameter constraints against the actual API documentation to ensure completeness.

[claude]

Pull Request Review: Add flux-2-max and gpt-image-1.5 models

Summary

This PR adds two new image generation models to the Celeste AI platform:

• flux-2-max: BFL FLUX.2 Max model with comprehensive parameter constraints
• gpt-image-1.5: OpenAI GPT Image 1.5 model

Overall, the implementation is clean and follows established patterns in the codebase. The code quality is good with minor suggestions below.

---

Code Quality & Best Practices ✅

Strengths:

1. Consistent with existing patterns: Both model definitions follow the exact structure used in other models in their respective files
2. Good documentation: Inline comments explain model-specific behaviors (e.g., "flux-2-max always upsamples prompts")
3. Clear parameter constraints: All parameters use appropriate constraint types (Dimensions, Choice, Range, Int)
4. Proper type annotations: Uses list[Model] type hints consistently

Minor Suggestions:

1. Version Bumping Strategy (packages/capabilities/image-generation/pyproject.toml:3)

The version was bumped from 0.3.3 to 0.3.5, skipping 0.3.4. While this works, it's worth confirming:

• Is there a reason for skipping 0.3.4?
• Consider documenting version bumping strategy in the repo if not already present

2. Streaming Parameter Inconsistency (packages/capabilities/image-generation/src/celeste_image_generation/providers/openai/models.py:63)

The gpt-image-1.5 model sets streaming=False, but the similar gpt-image-1 and gpt-image-1-mini models set streaming=True.

Question: Is this intentional? If GPT Image 1.5 doesn't support streaming, this is correct. Otherwise, this might be an oversight.

Compare:

• Line 33: gpt-image-1 has streaming=True
• Line 48: gpt-image-1-mini has streaming=True
• Line 63: gpt-image-1.5 has streaming=False ← Different

---

Potential Bugs or Issues ⚠️

flux-2-max Preset Validation

The flux-2-max model has extensive dimension presets (packages/capabilities/image-generation/src/celeste_image_generation/providers/bfl/models.py:19-28):

presets={
    "Square 1K": "1024x1024",
    "Square 2K": "2048x2048",
    "HD 16:9": "1920x1080",
    ...
}

Verification needed: Ensure all preset dimensions fall within the declared constraints:

• ✅ min_pixels=64*64 (4,096)
• ✅ max_pixels=2048*2048 (4,194,304)
• ✅ Aspect ratios between 9/21 (~0.429) and 21/9 (~2.333)

Quick check of presets:

• "Square 2K" (2048x2048) = 4,194,304 pixels ✅ (exactly at max)
• "Ultra-wide 21:9" (1920x832) = aspect ratio 2.307 ✅ (within bounds)
• "Portrait 9:21" (832x1920) = aspect ratio 0.433 ✅ (within bounds)

All presets appear valid! 👍

---

Performance Considerations 🚀

No performance concerns identified:

• Model definitions are static data structures loaded at import time
• No runtime overhead introduced
• Follows existing lightweight architecture (no vendor SDKs)

---

Security Concerns 🔒

No security issues detected:

• No user input handling in model definitions
• No API keys or sensitive data
• No code execution paths
• Pure data definitions using type-safe constraints

---

Test Coverage 📊

Current State:

Looking at the integration test (packages/capabilities/image-generation/tests/integration_tests/test_image_generation/test_generate.py:8-16), it tests:

• BFL: flux-2-pro ✅
• OpenAI: dall-e-2 ✅
• Google: imagen-4.0-fast-generate-001 ✅
• BytePlus: seedream-4-0-250828 ✅

Recommendation:

Consider adding integration tests for the new models:

@pytest.mark.parametrize(
    ("provider", "model", "parameters"),
    [
        # Existing tests...
        (Provider.BFL, "flux-2-max", {"aspect_ratio": "1024x1024", "safety_tolerance": 3}),
        (Provider.OPENAI, "gpt-image-1.5", {"aspect_ratio": "1024x1024", "quality": "medium"}),
    ],
)

Note: This is optional but would provide additional confidence, especially for parameter constraint validation.

---

Additional Observations

1. Documentation: The PR description includes helpful links to the model pages on withceleste.ai - good practice for reviewers to verify model specifications

2. Minimal diff: The PR touches only 4 files with focused changes - excellent for reducing merge conflicts and review burden

3. No breaking changes: This is a pure addition of new models with no modifications to existing functionality

---

Final Verdict: ✅ Approve with Minor Questions

This PR is well-structured and safe to merge. The only items to address before merging:

1. Clarify: Is the version jump from 0.3.3 → 0.3.5 intentional?
2. Verify: Is streaming=False correct for gpt-image-1.5 (vs other gpt-image models)?

Both are likely intentional, but worth confirming. Great work! 🎉